Skip to content

GitLab

Unverified Commit 2aba569a authored by Thomas Stocker's avatar Thomas Stocker Committed by GitHub
Browse files

Vulkan based on #9650 (#11835) 

* implement the vulkan C backend

* add support in gpu.go

* add support in gen_linux.sh

* it builds

* fix segfault

* fix compilation

* fix free memory monitor

* fix total memory monitor

* update gpu.go

* fix build

* fix check_perfmon len

* remove cap_get_bound check

* fix vulkan handle releasing

* fix build on federa 40

* fix vulkan on windows

* making amdgpu work on arm achitecutre with vulkan

* add x86_64 lines in VulkanGlobs and capLinuxGlobs

* add aarch64 lines in vulkanGlobs and capLinuxGlobs

* Fix variable name

* Add vulkan build patch from @jmorganca

* Sync vendored ggml to add Vulkan support

* Updated dockerfile

https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2660836871

Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Installing rocm library
Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* This version works well

built based on this: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2660836871

Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Applied 00-fix-vulkan-building.patch

Work done by McBane87 here: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2660836871

Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Fixed the "detached head" issues
Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Merged in the right direction
Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Merging the latest stable (#2)

* Applied 00-fix-vulkan-building.patch

* Implemented vulkan backend based on the work done by whyvl, Dts0, McBane87 and others

Tested on AMD Ryzen 7 8845HS w/ Radeon 780M Graphics with ROCm disabled

```
[GIN-debug] POST   /v1/chat/completions      --> github.com/ollama/ollama/server.(*Server).ChatHandler-fm (6 handlers)
[GIN-debug] POST   /v1/completions           --> github.com/ollama/ollama/server.(*Server).GenerateHandler-fm (6 handlers)
[GIN-debug] POST   /v1/embeddings            --> github.com/ollama/ollama/server.(*Server).EmbedHandler-fm (6 handlers)
[GIN-debug] GET    /v1/models                --> github.com/ollama/ollama/server.(*Server).ListHandler-fm (6 handlers)
[GIN-debug] GET    /v1/models/:model         --> github.com/ollama/ollama/server.(*Server).ShowHandler-fm (6 handlers)
time=2025-03-11T13:00:40.793Z level=INFO source=gpu.go:199 msg="vulkan: load libvulkan and libcap ok"
time=2025-03-11T13:00:40.877Z level=INFO source=gpu.go:421 msg="error looking up vulkan GPU memory" error="device is a CPU"
time=2025-03-11T13:00:40.878Z level=WARN source=amd_linux.go:443 msg="amdgpu detected, but no compatible rocm library found.  Either install rocm v6, or follow manual install instructions at https://github.com/ollama/ollama/blob/main/docs/linux.md#manual-install"
time=2025-03-11T13:00:40.878Z level=WARN source=amd_linux.go:348 msg="unable to verify rocm library: no suitable rocm found, falling back to CPU"
time=2025-03-11T13:00:40.879Z level=INFO source=types.go:137 msg="inference compute" id=0 library=vulkan variant="" compute=1.3 driver=1.3 name="AMD Radeon Graphics (RADV GFX1103_R1)" total="15.6 GiB" available="15.6 GiB"
```

```
 # ollama run phi4:14b
>>> /set verbose
Set 'verbose' mode.
>>> how's it going?
Hello! I'm here to help you with any questions or tasks you have. How can I assist you today? 😊



total duration:       3.341959745s
load duration:        18.165612ms
prompt eval count:    15 token(s)
prompt eval duration: 475ms
prompt eval rate:     31.58 tokens/s
eval count:           26 token(s)
eval duration:        2.846s
eval rate:            9.14 tokens/s
>>>
```

* This is no longer needed
Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Fixes SIGSEGV: segmentation violation running gemma3 models on ollama 0.6.0 #21

Patch provided by McBane87 on https://github.com/whyvl/ollama-vulkan/issues/21

Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Applied 04-disable-mmap-vulkan.patch

From: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2660836871

Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Pulled new upstream code for ggml-bulkan backend
Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Merged latest ollama 0.6.2 and nasrally's Flash Attention patches (#5)

* readme: add Ellama to list of community integrations (#9800)

* readme: add screenpipe to community integrations (#9786)

* Add support for ROCm gfx1151 (#9773)

* conditionally enable parallel pipelines

* sample: make mutations in transforms explicit (#9743)

* updated minP to use early exit making use of sorted tokens

* ml/backend/ggml: allocate memory with malloc when loading model (#9822)

* runner: remove cache prompt flag from ollama runner (#9826)

We do not need to bypass the prompt caching in the ollama runner yet, as
only embedding models needed to bypass the prompt caching. When embedding
models are implemented they can skip initializing this cache completely.

* ollamarunner: Check for minBatch of context space when shifting

Models can specify that a group of inputs need to be handled a single
batch. However, context shifting didn't respect this and could trigger
a break anyways. In this case, we should instead trigger a context
shift earlier so that it occurs before the grouped batch.

Note that there still some corner cases:
 - A long prompt that exceeds the context window can get truncated
   in the middle of an image. With the current models, this will
   result in the model not recognizing the image at all, which is
   pretty much the expected result with truncation.
 - The context window is set less than the minimum batch size. The
   only solution to this is to refuse to load the model with these
   settings. However, this can never occur with current models and
   default settings.

Since users are unlikely to run into these scenarios, fixing them is
left as a follow up.

* Applied latest patches from McBane87

See this for details: https://github.com/whyvl/ollama-vulkan/issues/7#issuecomment-2708820861

Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>

* Add ability to enable flash attention on vulkan (#4

)

* discover: add flash attention handling for vulkan
* envconfig: fix typo in config.go

As part of the process some code was refactored and I added a new field
FlashAttention to GpuInfo since the previous solution didn't allow for a
granular check via vulkan extensions. As a side effect, this now allows
for granular per-device FA support checking in other places

---------
Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>
Co-authored-by: default avatarzeo <108888572+zeozeozeo@users.noreply.github.com>
Co-authored-by: default avatarLouis Beaumont <louis.beaumont@gmail.com>
Co-authored-by: default avatarDaniel Hiltgen <dhiltgen@users.noreply.github.com>
Co-authored-by: default avatarMichael Yang <mxyng@pm.me>
Co-authored-by: default avatarParth Sareen <parth.sareen@ollama.com>
Co-authored-by: default avatarJeffrey Morgan <jmorganca@gmail.com>
Co-authored-by: default avatarBruce MacDonald <brucewmacdonald@gmail.com>
Co-authored-by: default avatarJesse Gross <jesse@ollama.com>
Co-authored-by: default avatarNikita <50599445+nasrally@users.noreply.github.com>

* Revert Readme changes

* Revert

* Revert changes in amd_linux.go

* Revert changes in amd_linux.go

* Remove flashattention setting gpu.go

* Revert whitespace changes in gpu.go

* Revert changes in transforms_test.go

* Revert changes in runner.go

* Revert changes in Makefile.sync

* Revert some unintented changes in Dockerfile

* Revert vulkan copy changes in Dockerfile

* Update Vulkan Code to de4c07f93783a1a96456a44dc16b9db538ee1618

* Fixed duplicate sync in ggml.go

* Revert changes in ggml.go

* Revert chnages in ggml.go

* enable falsh attention on vulkan

* revert remove parenthesis

* fixed flash attention logic enabling

* vk_check_flash_attention 0 means supported

* Update gpu.go

* Add vulkan to Windows Build script

* Remove commented out code

* Enable Vulkan Flash attention in FlashAttentionSupported

* Fix logging

* Update Vulkan backend to e54d41befcc1575f4c898c5ff4ef43970cead75f

* Removed libcap related code

libcap is not directly related to Vulkan and should be added by its own PR. It adds additional library dependencies for building and also requires users to run setcap or run ollama as root, which is not ideal for easy use

* Fix Unit Test (Add Vulkan Library)

* Add vulkan to TestHomogeneousGPUs
Test

* vulkan: get GPU ID (ollama v0.11.5)
Signed-off-by: default avatarXiaodong Ye <xiaodong.ye@mthreads.com>

* disable mmap for vulkan

* Reduce Changes remove TestHomogeneousGPUs (doesn't exist on master)

* Update vulkan version to the version used in llama.cpp

* rename gpu patch to correct number

* added Vulkan API to get correct Device UUID

current UUID from pipelineCacheUUID does not match CUDA

* Fix GPU ID Patch

* Remove Code not in llama.cpp

* modified UUID code inside ggml

* Fix Patch

* Copied minimal definition from vulkan header

* Fix compile error in Mac

Metal is preferred so we're disabling Vulkan for now

* Removed unused code

Fix linter error in CI

* Fix patches apply

* fixing lint error

* Removed unneeded function call

Somehow removing this call fixed the crashing when Vulkan header was removed

* added missing NL

* Fixed missing members in Vulkan header

also added zero clear for some structs

* Fixed wrong structure ID

* Fixed Vulkan header

More aligned with official header definition now

* buildvulkanAsSeperateFunction

* Vulkan on Windows Test

* temporarly comment out gate to run windows task

* use temporarly windows-latest for build

* Commenting out other presets to build vulkan

* reenable cpu

* commenting out error action stop

* temporarly commenting out rocm

* set vulkan path

* comment out cude for faster turnaround

* correct vulkan install

* correct vulkan silent install

* fixed install command

* revert debugging changes (vulkan builds on windows)

* revert windows-latest

* trying to build vulkan for linux

* temporarly disable cuda and rocm

* try again linux build

* fix version

* trying to fix

* trying again

* trying again

* fix version

* fixed vulkan-sdk name

* try again

* trying again

* try without version number

* try again

* add some more extra

* trying to use version 1.4.313

* revert debugging changes

* Filter out already supported gpus

* revert debug code

* Use runners for GPU discovery

This revamps how we discover GPUs in the system by leveraging the Ollama
runner.  This should eliminate inconsistency between our GPU discovery and the
runners capabilities at runtime, particularly for cases where we try to filter
out unsupported GPUs.  Now the runner does that implicitly based on the actual
device list.  In some cases free VRAM reporting can be unreliable which can
leaad to scheduling mistakes, so this also includes a patch to leverage more
reliable VRAM reporting libraries if available.

Automatic workarounds have been removed as only one GPU leveraged this, which
is now documented. This GPU will soon fall off the support matrix with the next
ROCm bump.

Additional cleanup of the scheduler and discovery packages can be done in the
future once we have switched on the new memory management code, and removed
support for the llama runner.

* timing info for runner

* WIP - wire up Vulkan with the new engine based discovery

Not a complete implementation - free VRAM is better, but not accurate on
windows

* fix - trust the library paths from discovery when starting runner

* fix index bug

* fix vulkan ids to be underlying

* fix - give bootstrapping more time on slow systems

* Test if Vulkan device is supported

* vk_check_flash_attention is not needed (coompat2 coopmapt and scalar implementation exist)

* Handle GGML_VK_VISIBLE_DEVICES

* ask for supported first

* win: fix CPU query buffer handling

Try in a short loop until we get the size right.

* test: harden integration tests for slow start

If the server takes a while to start up, block
tests from starting until it's online to avoid
setting large timeouts in individual test cases.

* gofumpt fix

* fix build

* merge fixes

* merge fixes

* fixed build

* merge fixes

* fixing build

* fixed build

* fixed formatting

* fixed build

* fix vulkan gpu id patch

* sync llama.cpp vulkan code

* update build windows script

* merge fixes

* fix format

* fixed vulkan casing

* handle igpu as gpu

* improve case

* print out unknown library

* rturn Vulkan for vulkan library

* Revert "rturn Vulkan for vulkan library"

This reverts commit 690461a12fd5e93295d174c97edefb2bc33285b1.

* fixed patch number

* return Library Name

* remvoe debug code

* return integrated in vulkan backend

* Return pci Properties

* update patch

* directly get pci proeprties without parsing

* workaround for filtering devices. Correct way is to have a LibraryPosition Parameter in the deviceInfo

* Revert "directly get pci proeprties without parsing"

This reverts commit 8e0624851f5ed7d9f74518f574dfb422e4dd4dc2.

* Set FilteredID for Environment Filtering

* ROCm Library is named ROCm

* revert changes in patch

* Create 0028-vulkan-pci-and-memory.patch

* vulkan memory patch

* casing fix

* Add more pci properties

* Added better memory management

* Added better memory managament

* fixed patch

* Fixed patch

* FilterID creation group by library

* filter out vulkan supported by other gpu

* fixing deviceid compare

* Vulkan Fix FA coopmat1 invalid array indexing

* Use everywhere the same Vulkan Version 1.4.321.1

* Remove unneeded patch

* vulkan update

* sync vulkan glsl files

* only use for vulkan the filteredid (numeric device number)

* simplify code

---------
Signed-off-by: default avatarVadim Grinco <vadim@grinco.eu>
Signed-off-by: default avatarXiaodong Ye <xiaodong.ye@mthreads.com>
Co-authored-by: default avatarpufferffish <github@bandersnatch.anonaddy.com>
Co-authored-by: KOISHI KOMEIJI FROM TOUHOU 11 <fuck>
Co-authored-by: default avatarDSLstandard <qgeneral35@gmail.com>
Co-authored-by: default avatarpufferffish <me@windtfw.com>
Co-authored-by: default avataryeongbba <yeongmo.lee@logpresso.com>
Co-authored-by: default avatartomaThomas <tomathomas@mailbox.org>
Co-authored-by: default avatarAntoine Viallon <antoine@lesviallon.fr>
Co-authored-by: default avatarVadim Grinco <vadim@grinco.eu>
Co-authored-by: default avatarzeo <108888572+zeozeozeo@users.noreply.github.com>
Co-authored-by: default avatarLouis Beaumont <louis.beaumont@gmail.com>
Co-authored-by: default avatarDaniel Hiltgen <dhiltgen@users.noreply.github.com>
Co-authored-by: default avatarMichael Yang <mxyng@pm.me>
Co-authored-by: default avatarParth Sareen <parth.sareen@ollama.com>
Co-authored-by: default avatarJeffrey Morgan <jmorganca@gmail.com>
Co-authored-by: default avatarBruce MacDonald <brucewmacdonald@gmail.com>
Co-authored-by: default avatarJesse Gross <jesse@ollama.com>
Co-authored-by: default avatarNikita <50599445+nasrally@users.noreply.github.com>
Co-authored-by: default avatarMasato Nakasaka <masato.nakasaka@intel.com>
Co-authored-by: default avatarXiaodong Ye <xiaodong.ye@mthreads.com>
Co-authored-by: default avatarDaniel Hiltgen <daniel@ollama.com>
parent fd8aa947
Hide whitespace changes
Inline Side-by-side
Showing with 2706 additions and 0 deletions
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "generic_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
#define FLT_MAX 3.402823466e+38F
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
shared FLOAT_TYPE tmpmax[BLOCK_SIZE];
shared uint tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
const uint col = gl_LocalInvocationID.x;
if (row >= p.KY) {
return;
}
A_TYPE amax = -FLT_MAX;
uint acol = col;
if (col < p.KX) {
amax = data_a[row*p.KX + col];
}
for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) {
A_TYPE val = data_a[row*p.KX + i];
if (val > amax) {
amax = val;
acol = i;
}
}
tmp[col] = acol;
tmpmax[col] = amax;
barrier();
[[unroll]] for (int s = int(BLOCK_SIZE) / 2; s > 0; s >>= 1) {
if (col < s && col + s < p.KX) {
if (tmpmax[col] < tmpmax[col + s]) {
tmpmax[col] = tmpmax[col + s];
tmp[col] = tmp[col + s];
}
}
barrier();
}
if (col == 0) {
data_d[row] = D_TYPE(tmp[0]);
}
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp 0 → 100644
View file @ 2aba569a
#version 450
#extension GL_EXT_control_flow_attributes : enable
#include "types.glsl"
layout(constant_id = 0) const int BLOCK_SIZE = 1024;
layout(constant_id = 1) const int BLOCK_SIZE_LOG2 = 10;
#define ASC 0
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) buffer D {int data_d[];};
layout (push_constant) uniform parameter {
uint ncols;
uint order;
} p;
shared int dst_row[BLOCK_SIZE];
shared A_TYPE a_sh[BLOCK_SIZE];
void swap(uint idx0, uint idx1) {
int tmp = dst_row[idx0];
dst_row[idx0] = dst_row[idx1];
dst_row[idx1] = tmp;
}
void argsort(bool needs_bounds_check) {
// bitonic sort
const int col = int(gl_LocalInvocationID.x);
const uint row = gl_WorkGroupID.y;
const uint row_offset = row * p.ncols;
// initialize indices
dst_row[col] = col;
a_sh[col] = data_a[row_offset + col];
barrier();
uint num_outer_loop_iters = BLOCK_SIZE_LOG2;
[[unroll]] for (uint k = 2, outer_idx = 0; outer_idx < num_outer_loop_iters; k *= 2, outer_idx++) {
uint num_inner_loop_iters = outer_idx + 1;
[[unroll]] for (uint j = k / 2, inner_idx = 0; inner_idx < num_inner_loop_iters; j /= 2, inner_idx++) {
const int ixj = int(col ^ j);
int idx_0 = (col & k) == 0 ? col : ixj;
int idx_1 = (col & k) == 0 ? ixj : col;
int sh_idx_0 = dst_row[idx_0];
int sh_idx_1 = dst_row[idx_1];
bool idx_0_oob = needs_bounds_check ? sh_idx_0 >= p.ncols : false;
bool idx_1_oob = needs_bounds_check ? sh_idx_1 >= p.ncols : false;
if ((idx_0_oob ||
(!idx_1_oob && a_sh[sh_idx_0] > a_sh[sh_idx_1])) && (ixj > col)) {
swap(idx_0, idx_1);
}
barrier();
}
}
if (col < p.ncols) {
if (p.order == ASC) {
data_d[row_offset + col] = dst_row[col];
} else {
data_d[row_offset + p.ncols - col - 1] = dst_row[col];
}
}
}
void main() {
if (p.ncols == BLOCK_SIZE) {
argsort(false);
} else {
argsort(true);
}
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "types.glsl"
#include "generic_unary_head.glsl"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
if (idx >= p.ne) {
return;
}
const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "types.glsl"
#include "generic_binary_head.glsl"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
const int dim = p.param3;
if (idx >= p.ne) {
return;
}
const uint i3 = idx / (p.ne22*p.ne21*p.ne20);
const uint i3_offset = i3 * p.ne22*p.ne21*p.ne20;
const uint i2 = (idx - i3_offset) / (p.ne21*p.ne20);
const uint i2_offset = i2*p.ne21*p.ne20;
const uint i1 = (idx - i3_offset - i2_offset) / p.ne20;
const uint i0 = idx - i3_offset - i2_offset - i1*p.ne20;
uint o[4] = {0, 0, 0, 0};
o[dim] = dim == 0 ? p.ne00 : (dim == 1 ? p.ne01 : (dim == 2 ? p.ne02 : p.ne03));
const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
const uint src1_idx = (i3 - o[3])*p.nb13 + (i2 - o[2])*p.nb12 + (i1 - o[1])*p.nb11 + (i0 - o[0])*p.nb10;
const uint dst_idx = i3*p.nb23 + i2*p.nb22 + i1*p.nb21 + i0*p.nb20;
const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
#ifndef OPTIMIZATION_ERROR_WORKAROUND
data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : data_b[get_boffset() + src1_idx]);
#else
if (is_src0) {
data_d[get_doffset() + dst_idx] = data_a[get_aoffset() + src0_idx];
} else {
data_d[get_doffset() + dst_idx] = data_b[get_boffset() + src1_idx];
}
#endif
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "types.glsl"
#include "generic_unary_head.glsl"
#extension GL_EXT_control_flow_attributes : require
const uint num_threads = 128;
layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
void main() {
uint idx = get_idx();
// num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
const uint num_iter = 4;
// fast path for when all four iterations are in-bounds
if (idx + (num_iter-1)*num_threads < p.ne) {
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
#if defined(DATA_D_BF16)
float f = float(data_a[get_aoffset() + idx]);
data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
#else
data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
#endif
idx += num_threads;
}
} else {
[[unroll]] for (uint i = 0; i < num_iter; ++i) {
if (idx >= p.ne) {
continue;
}
#if defined(DATA_D_BF16)
float f = float(data_a[get_aoffset() + idx]);
data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
#else
data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
#endif
idx += num_threads;
}
}
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_dw.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "types.glsl"
layout (push_constant) uniform parameter
{
uint ne;
uint batches;
uint channels;
uint dst_w;
uint dst_h;
uint src_w;
uint src_h;
uint knl_w;
uint knl_h;
int stride_x;
int stride_y;
int pad_x;
int pad_y;
int dilation_x;
int dilation_y;
} p;
layout (binding = 0) readonly buffer A {A_TYPE knl_data[];};
layout (binding = 1) readonly buffer B {B_TYPE src_data[];};
layout (binding = 2) writeonly buffer D {D_TYPE dst_data[];};
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
FLOAT_TYPE conv_2d_dw_whcn(uint idx) {
uint i0 = idx / p.dst_w;
uint dst_x = idx - i0 * p.dst_w;
uint i1 = i0 / p.dst_h;
uint dst_y = i0 - i1 * p.dst_h;
uint n = i1 / p.channels;
uint c = i1 - n * p.channels;
uint src_i = n * p.channels * p.src_h * p.src_w + c * p.src_h * p.src_w;
uint knl_i = c * p.knl_h * p.knl_w;
FLOAT_TYPE sum = 0.0;
for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) {
uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int
continue;
}
for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) {
uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int
continue;
}
FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * p.src_w + src_x]);
FLOAT_TYPE k = FLOAT_TYPE(knl_data[knl_i + knl_y * p.knl_w + knl_x]);
sum = fma(v, k, sum);
}
}
return sum;
}
FLOAT_TYPE conv_2d_dw_cwhn(uint idx) {
uint i0 = idx / p.channels;
uint c = idx - i0 * p.channels;
uint i1 = i0 / p.dst_w;
uint dst_x = i0 - i1 * p.dst_w;
uint n = i1 / p.dst_h;
uint dst_y = i1 - n * p.dst_h;
uint src_i = n * p.channels * p.src_h * p.src_w;
uint src_row = p.src_w * p.channels;
uint knl_row = p.knl_w * p.channels;
FLOAT_TYPE sum = 0.0;
for (uint knl_y = 0; knl_y < p.knl_h; ++knl_y) {
uint src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
if (src_y >= p.src_h) { // src_y < 0 will wrap to a large unsigned int
continue;
}
for (uint knl_x = 0; knl_x < p.knl_w; ++knl_x) {
uint src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
if (src_x >= p.src_w) { // src_x < 0 will wrap to a large unsigned int
continue;
}
FLOAT_TYPE v = FLOAT_TYPE(src_data[src_i + src_y * src_row + src_x * p.channels + c]);
FLOAT_TYPE k = FLOAT_TYPE(knl_data[ knl_y * knl_row + knl_x * p.channels + c]);
sum = fma(v, k, sum);
}
}
return sum;
}
void main() {
uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
if (idx >= p.ne) {
return;
}
FLOAT_TYPE result =
#ifdef WHCN
conv_2d_dw_whcn(idx);
#else
conv_2d_dw_cwhn(idx);
#endif
dst_data[idx] = D_TYPE(result);
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp 0 → 100644
View file @ 2aba569a
#version 450
#extension GL_EXT_control_flow_attributes : enable
#ifdef COOPMAT2
#extension GL_NV_cooperative_matrix2 : enable
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#extension GL_KHR_memory_scope_semantics : enable
#endif
#ifdef USE_COLLECTIVES
# extension GL_KHR_shader_subgroup_shuffle : enable
#endif
#include "types.glsl"
// shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
layout(binding = 0) readonly buffer A {
A_TYPE knl_data[];
}; // src0 - kernel: [KW, KH, Cin, Cout] for conv_2d, [KW, KH, Cout, Cin] for conv_transposed_2d
layout(binding = 1) readonly buffer B {
B_TYPE src_data[];
}; // src1 - input: [W, H, Cin, N] -- channel_first format
layout(binding = 2) writeonly buffer D {
D_TYPE dst_data[];
}; // dst - result: [OW, OH, Cout, N]
layout(push_constant) uniform parameter {
// I/O channels, batch size
uint32_t Cout;
uint32_t Cin;
uint32_t N;
// Tensor spatial sizes: kernel, input, output
uint32_t KW;
uint32_t KH;
uint32_t W;
uint32_t H;
uint32_t OW;
uint32_t OH;
// Parameters: stride, padding, dilation - 0=y, 1=x
uint32_t s0;
uint32_t s1;
uint32_t p0;
uint32_t p1;
uint32_t d0;
uint32_t d1;
// Strides in elements
uint32_t nb01;
uint32_t nb02;
uint32_t nb03;
uint32_t nb11;
uint32_t nb12;
uint32_t nb13;
uint32_t nb1;
uint32_t nb2;
uint32_t nb3;
// fastdiv helper values
uint32_t KWmp; uint32_t KWL;
uint32_t KWKHmp; uint32_t KWKHL;
uint32_t OWmp; uint32_t OWL;
uint32_t OWOHmp; uint32_t OWOHL;
#ifdef TRANSPOSE
uint32_t s0mp; uint32_t s0L;
uint32_t s1mp; uint32_t s1L;
#endif
}
p;
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
// Blocktile sizes
layout(constant_id = 1) const uint BS_K = 128;
layout(constant_id = 2) const uint BS_CRS = 16;
layout(constant_id = 3) const uint BS_NPQ = 128;
// Thread-tile sizes
layout(constant_id = 4) const uint TS_K = 8;
layout(constant_id = 5) const uint use_collectives = 1;
layout(constant_id = 6) const uint SHMEM_PAD = 4;
uint32_t tid = gl_LocalInvocationID.x;
const uint32_t WG_SIZE = gl_WorkGroupSize.x;
uint splitWork(uint work_size, uint block_size) {
return (block_size + work_size - 1) / block_size;
}
uint32_t K = p.Cout;
uint32_t CRS = p.Cin * p.KH * p.KW;
uint32_t NPQ = p.N * p.OH * p.OW;
uint32_t n_elems_out = K * NPQ;
// Number of blocktiles per input
uint32_t NB_CRS = splitWork(CRS, BS_CRS);
#ifdef COOPMAT2
#define SHMEM_TYPE float16_t
#else
#define SHMEM_TYPE float
#endif
const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
const uint32_t Ash_numel = BS_K * BS_CRS;
const uint32_t Bsh_numel = BS_CRS * BS_NPQ;
const uint32_t Ash_len = BS_K * Ash_stride;
const uint32_t Bsh_len = BS_CRS * Bsh_stride;
shared SHMEM_TYPE Ash[Ash_len]; // K x CRS
shared SHMEM_TYPE Bsh[Bsh_len]; // CRS x NPQ
// Threadtile sizes
const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
// Number of threadtiles per blocktile
const uint32_t NT_K = BS_K / TS_K;
const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
/*
Compute
KxCRS @ CRSxNPQ = K x NPQ
K=Cout
C=Cin
R,S=KH,KW
P,Q=OH,OW
*/
uint32_t B_idx_K = gl_WorkGroupID.x;
uint32_t B_idx_NPQ = gl_WorkGroupID.y;
uint32_t T_y = tid / NT_NPQ;
uint32_t T_x = tid % NT_NPQ;
uint32_t Ar = tid / BS_CRS;
uint32_t Ac = tid % BS_CRS;
const uint32_t ArpWg = WG_SIZE / BS_CRS;
uint32_t Br = tid / BS_NPQ;
uint32_t Bc = tid % BS_NPQ;
const uint32_t BrpWg = WG_SIZE / BS_NPQ;
// see init_fastdiv_values in ggml-vulkan.cpp
uint fastdiv(uint n, uint mp, uint L) {
uint msbs, lsbs;
// msbs = mulhi(n, mp)
umulExtended(n, mp, msbs, lsbs);
return (msbs + n) >> L;
}
#ifdef COOPMAT2
#define ACC_TYPE float16_t
ACC_TYPE perElemOpStore(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem)
{
uint32_t K_idx = B_idx_K * BS_K + r;
uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + c;
uint32_t N_idx = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
uint32_t OH_idx = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
uint32_t OW_idx = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
if (K_idx < K && NPQ_idx < NPQ) {
dst_data[dst_idx] = D_TYPE(elem);
}
return elem;
}
#endif
void main() {
#ifdef COOPMAT2
coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator> matC;
matC = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator>(0.0);
#else
float regC[TS_K][TS_NPQ];
for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
regC[T_ly][T_lx] = 0.0;
}
}
#endif
/* Advance block in CRS dim */
for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
uint32_t CRS_idx_a;
uint32_t Cin_idx_a;
uint32_t KH_idx_a;
uint32_t KW_idx_a;
#ifdef USE_COLLECTIVES
uint32_t cached_CRS_idx;
uint32_t cached_Cin_idx;
uint32_t cached_KH_idx;
uint32_t cached_KW_idx;
if (use_collectives == 1) {
cached_CRS_idx = B_idx_CRS * BS_CRS + gl_SubgroupInvocationID;
cached_Cin_idx = fastdiv(cached_CRS_idx, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
uint32_t cached_CRS_remainder = (cached_CRS_idx - cached_Cin_idx * p.KW * p.KH);
cached_KH_idx = fastdiv(cached_CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
cached_KW_idx = cached_CRS_remainder - cached_KH_idx * p.KW;
CRS_idx_a = subgroupShuffle(cached_CRS_idx, Ac);
Cin_idx_a = subgroupShuffle(cached_Cin_idx, Ac);
KH_idx_a = subgroupShuffle(cached_KH_idx, Ac);
KW_idx_a = subgroupShuffle(cached_KW_idx, Ac);
} else {
CRS_idx_a = B_idx_CRS * BS_CRS + Ac; // Global CRS_idx_a (column index of A)
Cin_idx_a = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
uint32_t CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
KH_idx_a = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
KW_idx_a = CRS_remainder - KH_idx_a * p.KW;
}
#else
CRS_idx_a = B_idx_CRS * BS_CRS + Ac; // Global CRS_idx_a (column index of A)
Cin_idx_a = fastdiv(CRS_idx_a, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH); / (p.KW * p.KH);
CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
KH_idx_a = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
KW_idx_a = CRS_remainder - KH_idx_a * p.KW;
#endif
/* Load kernel to A_block: (BS_K x BS_CRS)*/
for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
uint32_t B_ly = r_offset + Ar;
uint32_t B_lx = Ac;
uint32_t K_idx = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
#ifdef TRANSPOSE
uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + K_idx * p.nb02 + Cin_idx_a * p.nb03, K * CRS - 1);
#else
uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + Cin_idx_a * p.nb02 + K_idx * p.nb03, K * CRS - 1);
#endif
float val = knl_data[knl_idx];
if (K_idx >= K || CRS_idx_a >= CRS) {
val = 0.0;
}
Ash[B_ly * Ash_stride + B_lx] = SHMEM_TYPE(val);
}
/* Load input to B_block: (BS_CRS x BS_NPQ) */
UNROLL for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
uint32_t B_ly = r_offset + Br; /* Row index of B block */
uint32_t B_lx = Bc;
uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
uint32_t N_idx = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
uint32_t NPQ_remainder = NPQ_idx - N_idx * p.OH * p.OW;
uint32_t OH_idx = fastdiv(NPQ_remainder, p.OWmp, p.OWL); // divide by p.OW;
uint32_t OW_idx = NPQ_remainder - OH_idx * p.OW;
uint32_t CRS_idx_b;
uint32_t Cin_idx_b;
uint32_t KH_idx_b;
uint32_t KW_idx_b;
#ifdef USE_COLLECTIVES
if (use_collectives == 1) {
CRS_idx_b = subgroupShuffle(cached_CRS_idx, r_offset + Br);
Cin_idx_b = subgroupShuffle(cached_Cin_idx, r_offset + Br);
KH_idx_b = subgroupShuffle(cached_KH_idx, r_offset + Br);
KW_idx_b = subgroupShuffle(cached_KW_idx, r_offset + Br);
} else {
CRS_idx_b = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
Cin_idx_b = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
KH_idx_b = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
KW_idx_b = CRS_remainder - KH_idx_b * p.KW;
}
#else
CRS_idx_b = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
Cin_idx_b = fastdiv(CRS_idx_b, p.KWKHmp, p.KWKHL); // divide by (p.KW * p.KH);
uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
KH_idx_b = fastdiv(CRS_remainder, p.KWmp, p.KWL); // divide by p.KW;
KW_idx_b = CRS_remainder - KH_idx_b * p.KW;
#endif
#ifdef TRANSPOSE
uint32_t H_idx_x_s1 = OH_idx - KH_idx_b * p.d1 + p.p1;
uint32_t W_idx_x_s0 = OW_idx - KW_idx_b * p.d0 + p.p0;
uint32_t H_idx = fastdiv(H_idx_x_s1, p.s1mp, p.s1L);
uint32_t W_idx = fastdiv(W_idx_x_s0, p.s0mp, p.s0L);
#else
uint32_t H_idx = OH_idx * p.s1 + KH_idx_b * p.d1 - p.p1;
uint32_t W_idx = OW_idx * p.s0 + KW_idx_b * p.d0 - p.p0;
#endif
uint32_t src_idx =
min(max(W_idx + H_idx * p.nb11 + Cin_idx_b * p.nb12 + N_idx * p.nb13, 0), p.Cin * p.N * p.W * p.H - 1);
float val = src_data[src_idx];
if (CRS_idx_b >= CRS || NPQ_idx >= NPQ
|| H_idx >= p.H || W_idx >= p.W // Lower bound checks aren't necessary. (idx >= 0x80000000 for such case)
#ifdef TRANSPOSE
|| (H_idx_x_s1 - H_idx * p.s1 != 0) || (W_idx_x_s0 - W_idx * p.s0 != 0)
#endif
) {
val = 0.0;
}
Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);
}
barrier();
#ifdef COOPMAT2
coopmat<float16_t, gl_ScopeWorkgroup, BS_K, BS_CRS, gl_MatrixUseA> matA;
coopmat<float16_t, gl_ScopeWorkgroup, BS_CRS, BS_NPQ, gl_MatrixUseB> matB;
coopMatLoad(matA, Ash, 0, Ash_stride, gl_CooperativeMatrixLayoutRowMajor);
coopMatLoad(matB, Bsh, 0, Bsh_stride, gl_CooperativeMatrixLayoutRowMajor);
matC = coopMatMulAdd(matA, matB, matC);
#else
if (T_y * TS_K < K) {
UNROLL for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
float regA[TS_K];
float regB[TS_NPQ];
for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
}
for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
}
for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
}
}
}
}
#endif
barrier();
}
/* Save C* */
#ifdef COOPMAT2
coopMatPerElementNV(matC, matC, perElemOpStore);
#else
if (T_y * TS_K < K) {
for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
uint32_t K_idx = B_idx_K * BS_K + T_y * TS_K + T_ly;
uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
uint32_t N_idx = fastdiv(NPQ_idx, p.OWOHmp, p.OWOHL); // divide by p.OH * p.OW;
uint32_t OH_idx = fastdiv(NPQ_idx - N_idx * p.OH * p.OW, p.OWmp, p.OWL); // divide by p.OW;
uint32_t OW_idx = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
if (K_idx < K && NPQ_idx < NPQ) {
dst_data[dst_idx] = regC[T_ly][T_lx];
}
}
}
}
#endif
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/conv_transpose_1d.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "types.glsl"
layout (binding = 0) readonly buffer A {A_TYPE data_a[];}; // src0 - kernel: [K, Cout, Cin]
layout (binding = 1) readonly buffer B {B_TYPE data_b[];}; // src1 - input: [L, Cin]
layout (binding = 2) writeonly buffer D {D_TYPE data_d[];}; // dst - result [KL, Cout]
layout(local_size_x = 128 , local_size_y = 1, local_size_z = 1) in;
layout (push_constant) uniform parameter {
uint32_t Cout;
uint32_t Cin;
uint32_t K;
uint32_t L;
uint32_t KL;
uint32_t nb01;
uint32_t nb02;
uint32_t nb11;
uint32_t nb1;
int32_t s0;
} p;
uint32_t Cout_idx = gl_WorkGroupID.x;
const uint32_t bs = gl_WorkGroupSize.x;
uint32_t tid = gl_LocalInvocationID.x;
// Code is more straightforward if we assume it is bs*s0+K instead of (bs-1)*s0+K.
uint32_t tmp_len = bs*p.s0+p.K;
shared D_TYPE tmp[4096];
uint splitWork(uint workSize){
return (bs + workSize -1) / bs;
}
void main(){
for(uint32_t i = 0; i < splitWork(tmp_len); i++){
uint32_t idx = i*bs+tid;
if(idx < tmp_len){
tmp[idx] = 0.0;
}
}
uint32_t L_blocks = splitWork(p.L);
for(uint32_t L_block_id = 0; L_block_id < L_blocks; L_block_id++){
if(L_block_id > 0){
barrier();
// Shift values in tmp to the current processing window
for(int i = 0; i < splitWork(tmp_len); i++){
uint32_t idx = i*bs+tid;
if(idx >= bs*p.s0 && idx < tmp_len){
tmp[idx-bs*p.s0] = tmp[idx];
tmp[idx] = 0.0;
}else if(idx >= p.K && idx < bs*p.s0){
tmp[idx] = 0.0;
}
}
}
barrier();
// Save contributions of the block to tmp
uint32_t L_idx = L_block_id*bs + tid;
for(uint32_t K_idx = 0; K_idx < p.K; K_idx++){
D_TYPE dp = 0.0;
for(uint32_t Cin_idx = 0; Cin_idx < p.Cin; Cin_idx++){
A_TYPE elemKrn = data_a[K_idx + Cout_idx * p.nb01 + Cin_idx * p.nb02];
if(L_idx < p.L){
B_TYPE elemInp = data_b[L_idx + Cin_idx*p.nb11];
dp = fma(elemKrn, elemInp, dp);
}
}
tmp[tid*p.s0 + K_idx] += dp;
barrier();
}
// Save the computed values except the last block that can have different size
uint32_t KLb_idx = L_block_id*bs*p.s0;
if(L_block_id < L_blocks-1){
for(uint32_t s0_idx = 0; s0_idx < p.s0; s0_idx++){
uint32_t sh_idx = p.s0*tid+s0_idx;
uint32_t KL_idx = KLb_idx+sh_idx;
if(KL_idx < p.KL){
data_d[KL_idx + Cout_idx*p.nb1] = tmp[sh_idx];
}
}
}
}
for(uint32_t i = 0; i < splitWork(tmp_len); i++){
uint32_t idx = i*bs+tid;
uint32_t KL_idx = (L_blocks-1)*bs*p.s0+idx;
if(KL_idx < p.KL){
data_d[KL_idx + Cout_idx*p.nb1] = tmp[idx];
}
}
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "types.glsl"
#include "generic_unary_head.glsl"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
if (idx >= p.ne) {
return;
}
#if defined(DATA_D_BF16)
float f = float(data_a[get_aoffset() + src0_idx(idx)]);
data_d[get_doffset() + dst_idx(idx)] = D_TYPE(fp32_to_bf16(f));
#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
#else
data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)];
#endif
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "types.glsl"
#include "generic_unary_head.glsl"
#include "dequant_funcs.glsl"
#if defined(DATA_A_IQ4_NL) || defined(DATA_A_MXFP4)
// 16 invocations needed for init_iq_shmem
layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
#else
layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
#endif
void main() {
#ifdef NEEDS_INIT_IQ_SHMEM
init_iq_shmem(gl_WorkGroupSize);
if (gl_LocalInvocationIndex.x != 0) {
return;
}
#endif
const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
if (idx >= p.ne) {
return;
}
uint dst_idx = get_doffset() + dst_idx(idx);
uint src_idx = src0_idx_quant(idx, QUANT_K);
const uint a_offset = 0;
const uint ib = src_idx;
const vec2 dm = get_dm(ib, a_offset);
[[unroll]] for (int j = 0; j < QUANT_K; j += 4) {
vec4 v = dequantize4(ib, j / QUANT_R, a_offset);
v = v * dm.x + vec4(dm.y);
#if QUANT_R == 2
data_d[dst_idx + j/2 + 0] = v[0];
data_d[dst_idx + j/2 + QUANT_K/2 + 0] = v[1];
data_d[dst_idx + j/2 + 1] = v[2];
data_d[dst_idx + j/2 + QUANT_K/2 + 1] = v[3];
#else
data_d[dst_idx + j + 0] = v[0];
data_d[dst_idx + j + 1] = v[1];
data_d[dst_idx + j + 2] = v[2];
data_d[dst_idx + j + 3] = v[3];
#endif
}
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "rte.glsl"
#include "types.glsl"
#if defined(SET_ROWS) && QUANT_K == 1
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
const uint BLOCK_SIZE = 512;
#else
layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
const uint BLOCK_SIZE = 32;
#endif
layout (binding = 0) readonly buffer S {float data_s[];};
#if defined(SET_ROWS)
#include "generic_binary_head.glsl"
layout (binding = 1) readonly buffer C {B_TYPE data_i[];};
layout (binding = 2) writeonly buffer Q {A_TYPE data_q[];};
#if B_SIZE == 64
#define DATA_I_SWIZZLE .x
#else
#define DATA_I_SWIZZLE
#endif
#else
#include "generic_unary_head.glsl"
layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];};
#endif
#if defined(DATA_A_Q4_0)
void quantize(uint dst_idx, uint src_idx)
{
float amax = 0.0;
float vmax = 0.0;
[[unroll]] for (int j = 0; j < QUANT_K_Q4_0; ++j) {
const float v = data_s[src_idx + j];
if (amax < abs(v)) {
amax = abs(v);
vmax = v;
}
}
const float d = vmax / -8;
const float id = (d != 0.0) ? 1.0/d : 0.0;
data_q[dst_idx].d = float16_t(d);
[[unroll]] for (int j = 0; j < QUANT_K_Q4_0/2; ++j) {
const float x0 = data_s[src_idx + 0 + j]*id;
const float x1 = data_s[src_idx + QUANT_K_Q4_0/2 + j]*id;
const uint xi0 = min(15, int(x0 + 8.5));
const uint xi1 = min(15, int(x1 + 8.5));
data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4));
}
}
#endif
#if defined(DATA_A_Q4_1)
void quantize(uint dst_idx, uint src_idx)
{
float vmin = 1.0/0.0;
float vmax = -vmin;
[[unroll]] for (int j = 0; j < QUANT_K_Q4_1; ++j) {
const float v = data_s[src_idx + j];
if (v < vmin) vmin = v;
if (v > vmax) vmax = v;
}
const float d = (vmax - vmin) / ((1 << 4) - 1);
const float id = (d != 0.0) ? 1.0/d : 0.0;
data_q[dst_idx].d = float16_t(d);
data_q[dst_idx].m = float16_t(vmin);
[[unroll]] for (int j = 0; j < QUANT_K_Q4_1/2; ++j) {
const float x0 = (data_s[src_idx + 0 + j] - vmin)*id;
const float x1 = (data_s[src_idx + QUANT_K_Q4_1/2 + j] - vmin)*id;
const uint xi0 = min(15, int(x0 + 0.5));
const uint xi1 = min(15, int(x1 + 0.5));
data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4));
}
}
#endif
#if defined(DATA_A_Q5_0)
void quantize(uint dst_idx, uint src_idx)
{
float amax = 0.0;
float vmax = 0.0;
[[unroll]] for (int j = 0; j < QUANT_K_Q5_0; ++j) {
const float v = data_s[src_idx + j];
if (amax < abs(v)) {
amax = abs(v);
vmax = v;
}
}
const float d = vmax / -16;
const float id = (d != 0.0) ? 1.0/d : 0.0;
data_q[dst_idx].d = float16_t(d);
uint32_t qh = 0;
[[unroll]] for (int j = 0; j < QUANT_K_Q5_0/2; ++j) {
const float x0 = data_s[src_idx + 0 + j]*id;
const float x1 = data_s[src_idx + QUANT_K_Q5_0/2 + j]*id;
const uint xi0 = min(31, int(x0 + 16.5));
const uint xi1 = min(31, int(x1 + 16.5));
data_q[dst_idx].qs[j] = uint8_t((xi0 & 0xf) | ((xi1 & 0xf) << 4));
qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
qh |= ((xi1 & 0x10u) >> 4) << (j + QUANT_K_Q5_0/2);
}
data_q[dst_idx].qh[0] = uint16_t(qh & 0xFFFF);
data_q[dst_idx].qh[1] = uint16_t(qh >> 16);
}
#endif
#if defined(DATA_A_Q5_1)
void quantize(uint dst_idx, uint src_idx)
{
float min = data_s[src_idx + 0];
float max = min;
[[unroll]] for (int j = 1; j < QUANT_K_Q5_1; ++j) {
const float v = data_s[src_idx + j];
min = v < min ? v : min;
max = v > max ? v : max;
}
const float d = (max - min) / 31;
const float id = (d != 0) ? 1.0/d : 0.0;
data_q[dst_idx].d = float16_t(d);
data_q[dst_idx].m = float16_t(min);
uint32_t qh = 0;
[[unroll]] for (int j = 0; j < QUANT_K_Q5_1/2; ++j) {
const float x0 = (data_s[src_idx + 0 + j] - min)*id;
const float x1 = (data_s[src_idx + QUANT_K_Q5_1/2 + j] - min)*id;
const uint xi0 = uint(x0 + 0.5);
const uint xi1 = uint(x1 + 0.5);
data_q[dst_idx].qs[j] = uint8_t((xi0 & 0xf) | ((xi1 & 0xf) << 4));
qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
qh |= ((xi1 & 0x10u) >> 4) << (j + QUANT_K_Q5_1/2);
}
data_q[dst_idx].qh = qh;
}
#endif
#if defined(DATA_A_Q8_0)
void quantize(uint dst_idx, uint src_idx)
{
float amax = 0.0; // absolute max
[[unroll]] for (int j = 0; j < QUANT_K_Q8_0; j++) {
const float v = data_s[src_idx + j];
amax = max(amax, abs(v));
}
const float d = amax / ((1 << 7) - 1);
const float id = (d != 0.0) ? 1.0/d : 0.0;
data_q[dst_idx].d = float16_t(d);
[[unroll]] for (int j = 0; j < QUANT_K_Q8_0; ++j) {
const float x0 = data_s[src_idx + j]*id;
data_q[dst_idx].qs[j] = int8_t(round(x0));
}
}
#endif
#if defined(DATA_A_IQ4_NL)
uint best_index(float x) {
if (x <= kvalues_iq4nl[0]) return 0;
if (x >= kvalues_iq4nl[15]) return 15;
int ml = 0, mu = 15;
while (mu-ml > 1) {
int mav = (ml+mu)/2;
if (x < kvalues_iq4nl[mav]) mu = mav; else ml = mav;
}
return x - kvalues_iq4nl[mu-1] < kvalues_iq4nl[mu] - x ? mu-1 : mu;
}
void quantize(uint dst_idx, uint src_idx)
{
float amax = 0.0;
float vmax = 0.0;
[[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL; ++j) {
const float v = data_s[src_idx + j];
if (amax < abs(v)) {
amax = abs(v);
vmax = v;
}
}
float d = vmax / kvalues_iq4nl[0];
const float id = (d != 0.0) ? 1.0/d : 0.0;
float sumqx = 0, sumq2 = 0;
[[unroll]] for (int j = 0; j < QUANT_K_IQ4_NL/2; ++j) {
const float x0 = data_s[src_idx + 0 + j]*id;
const float x1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*id;
const uint xi0 = best_index(x0);
const uint xi1 = best_index(x1);
data_q[dst_idx].qs[j] = uint8_t(xi0 | (xi1 << 4));
const float v0 = kvalues_iq4nl[xi0];
const float v1 = kvalues_iq4nl[xi1];
const float w0 = data_s[src_idx + 0 + j]*data_s[src_idx + 0 + j];
const float w1 = data_s[src_idx + QUANT_K_IQ4_NL/2 + j]*data_s[src_idx + QUANT_K_IQ4_NL/2 + j];
sumqx += w0*v0*data_s[src_idx + j] + w1*v1*data_s[src_idx + QUANT_K_IQ4_NL/2 + j];
sumq2 += w0*v0*v0 + w1*v1*v1;
}
data_q[dst_idx].d = float16_t(sumq2 > 0 ? sumqx/sumq2 : d);
}
#endif
#if defined(DATA_A_F32) || defined(DATA_A_F16)
void quantize(uint dst_idx, uint src_idx)
{
data_q[dst_idx] = A_TYPE(data_s[src_idx]);
}
#endif
#if defined(DATA_A_BF16)
void quantize(uint dst_idx, uint src_idx)
{
data_q[dst_idx] = A_TYPE(fp32_to_bf16(data_s[src_idx]));
}
#endif
#if defined(SET_ROWS)
void main() {
#ifdef NEEDS_INIT_IQ_SHMEM
init_iq_shmem(gl_WorkGroupSize);
#endif
const uint idx = ((gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x) * BLOCK_SIZE + gl_LocalInvocationID.x) * QUANT_K;
if (idx >= p.ne) {
return;
}
uint i00, i01, i02, i03;
get_indices(idx, i00, i01, i02, i03);
uint i12 = fastmod(i03, p.ne12);
uint i11 = fastmod(i02, p.ne11);
uint i10 = i01;
uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()] DATA_I_SWIZZLE;
uint src0_idx = src0_idx(i00, i01, i02, i03) + get_aoffset();
uint dst_idx = dst_idx(i00 / QUANT_K, i1, i02, i03) + get_doffset();
quantize(dst_idx, src0_idx);
}
#else
void main() {
#ifdef NEEDS_INIT_IQ_SHMEM
init_iq_shmem(gl_WorkGroupSize);
#endif
const uint idx = (gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x) * QUANT_K;
if (idx >= p.ne) {
return;
}
uint dst_idx = dst_idx_quant(idx, QUANT_K);
uint src_idx = get_aoffset() + src0_idx(idx);
quantize(dst_idx, src_idx);
}
#endif
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "types.glsl"
#include "generic_unary_head.glsl"
layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
void main() {
const uint idx = get_idx();
if (idx >= p.ne) {
return;
}
const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
data_d[get_doffset() + dst_idx(idx)] = D_TYPE(cos(val));
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/count_equal.comp 0 → 100644
View file @ 2aba569a
#version 450
#extension GL_EXT_control_flow_attributes : enable
#include "types.glsl"
#include "generic_head.glsl"
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
layout (binding = 2) buffer D {D_TYPE data_d[];};
const uint CHUNK_SIZE = 512;
void main() {
const uint base = gl_WorkGroupID.x * CHUNK_SIZE;
const uint col = gl_LocalInvocationID.x;
uint count = 0;
[[unroll]]
for (uint i = 0; i < CHUNK_SIZE; i += gl_WorkGroupSize.x) {
const uint idx = base + i + col;
if (idx >= p.KX) {
break;
}
count += uint(data_a[idx] == data_b[idx]);
}
atomicAdd(data_d[0], D_TYPE(count));
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "dequant_head.glsl"
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {float data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
const uint i = gl_GlobalInvocationID.x * 16;
if (i >= p.nel) {
return;
}
[[unroll]] for (uint l = 0; l < 16; l++) {
data_b[i + l] = D_TYPE(data_a[i + l]);
}
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl 0 → 100644
View file @ 2aba569a
#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
#endif
#include "types.glsl"
#if defined(A_TYPE_PACKED16)
layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
#endif
#if defined(A_TYPE_PACKED32)
layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
#endif
#if defined(DATA_A_F32)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
}
#endif
#if defined(DATA_A_F16)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
}
#endif
#if defined(DATA_A_BF16)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
return vec2(bf16_to_fp32(data_a[a_offset + ib]), bf16_to_fp32(data_a[a_offset + ib + 1]));
}
#endif
#if defined(DATA_A_Q4_0)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return (vec2(vui & 0xF, vui >> 4) - 8.0f);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12) - 8.0f);
}
#endif
#if defined(DATA_A_Q4_1)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return vec2(vui & 0xF, vui >> 4);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12);
}
#endif
#if defined(DATA_A_Q5_0)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0];
const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
return (vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) - 16.0f);
}
#endif
#if defined(DATA_A_Q5_1)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint uint_qh = data_a[a_offset + ib].qh;
const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint uint_qh = data_a_packed16[a_offset + ib].qh;
const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
return vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y);
}
#endif
#if defined(DATA_A_Q8_0)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1]));
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const i8vec2 v0 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2])).xy; // vec4 used due to #12147
const i8vec2 v1 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2 + 1])).xy;
return vec4(v0.x, v0.y, v1.x, v1.y);
}
#endif
#if defined(DATA_A_IQ1_S)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint ib32 = iqs / 32;
const uint ib8 = iqs / 8;
const int i8 = int(iqs % 8);
const uint qh = data_a[a_offset + ib].qh[ib32];
const uint qs = data_a[a_offset + ib].qs[ib8];
const float dl = float(2 * bitfieldExtract(qh, 12, 3) + 1);
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
const uint idxhi = bitfieldExtract(qh, 3 * int(ib8 & 3), 3);
const int16_t grid = int16_t(iq1s_grid[qs | (idxhi << 8)]);
// Signed bitfield extract.
const ivec2 gvec = ivec2(
bitfieldExtract(grid, 2 * (i8), 2),
bitfieldExtract(grid, 2 * (i8 + 1), 2)
);
return dl * (vec2(gvec) + delta);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint ib32 = iqs / 32;
const uint ib8 = iqs / 8;
const int i8 = int(iqs % 8);
const uint qh = data_a[a_offset + ib].qh[ib32];
const uint qs = data_a[a_offset + ib].qs[ib8];
const float dl = 2 * bitfieldExtract(qh, 12, 3) + 1;
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]);
// Signed bitfield extract.
const ivec4 gvec = ivec4(
bitfieldExtract(grid, 2 * (i8), 2),
bitfieldExtract(grid, 2 * (i8 + 1), 2),
bitfieldExtract(grid, 2 * (i8 + 2), 2),
bitfieldExtract(grid, 2 * (i8 + 3), 2)
);
return dl * (vec4(gvec) + delta);
}
#endif
#if defined(DATA_A_IQ1_M)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint ib8 = iqs / 8;
const uint ib16 = iqs / 16;
const int i8 = int(iqs % 8);
const uint sc = data_a[a_offset + ib].scales[iqs / 64];
const uint qs = data_a[a_offset + ib].qs[ib8];
const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1));
const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
// Signed bitfield extract.
const ivec2 gvec = ivec2(
bitfieldExtract(grid, 2 * (i8), 2),
bitfieldExtract(grid, 2 * (i8 + 1), 2)
);
return dl * (vec2(gvec) + delta);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint ib8 = iqs / 8;
const uint ib16 = iqs / 16;
const int i8 = int(iqs % 8);
const uint sc = data_a[a_offset + ib].scales[iqs / 64];
const uint qs = data_a[a_offset + ib].qs[ib8];
const uint qh = data_a[a_offset + ib].qh[ib16] >> (4 * (ib8 & 1));
const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
// Signed bitfield extract.
const ivec4 gvec = ivec4(
bitfieldExtract(grid, 2 * (i8), 2),
bitfieldExtract(grid, 2 * (i8 + 1), 2),
bitfieldExtract(grid, 2 * (i8 + 2), 2),
bitfieldExtract(grid, 2 * (i8 + 3), 2)
);
return dl * (vec4(gvec) + delta);
}
#endif
#if defined(DATA_A_IQ2_XXS)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint ib32 = iqs / 32;
const uint ib8 = (iqs / 8) % 4;
const uint qs = data_a[a_offset + ib].qs[8 * ib32 + ib8];
// Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale)
const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[4 * ib32 + 2],
data_a_packed16[a_offset + ib].qs[4 * ib32 + 3]));
const float db = 0.25 * (0.5 + (signs >> 28));
const uint sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
// Add parity bit
const uint sign8 = sign7 | (bitCount(sign7) << 7);
const uint sign = sign8 >> (iqs % 8);
const u8vec4 grid = unpack8(iq2xxs_grid[qs][(iqs % 8) / 4] >> (8 * (iqs % 4)));
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
return db * vec2(
grid.x * (sign0 ? -1.0 : 1.0),
grid.y * (sign1 ? -1.0 : 1.0)
);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint ib32 = iqs / 32;
const uint ib8 = (iqs / 8) % 4;
const uint qs = data_a[a_offset + ib].qs[8 * ib32 + ib8];
// Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale)
const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[4 * ib32 + 2],
data_a_packed16[a_offset + ib].qs[4 * ib32 + 3]));
const float db = 0.25 * (0.5 + (signs >> 28));
const uint sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
// Add parity bit
const uint sign8 = sign7 | (bitCount(sign7) << 7);
const uint sign = sign8 >> (iqs % 8);
const u8vec4 grid = unpack8(iq2xxs_grid[qs][(iqs % 8) / 4] >> (8 * (iqs % 4)));
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
bool sign2 = (sign & 4) != 0;
bool sign3 = (sign & 8) != 0;
return db * vec4(
grid.x * (sign0 ? -1.0 : 1.0),
grid.y * (sign1 ? -1.0 : 1.0),
grid.z * (sign2 ? -1.0 : 1.0),
grid.w * (sign3 ? -1.0 : 1.0)
);
}
#endif
#if defined(DATA_A_IQ2_XS)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint scale = (data_a[a_offset + ib].scales[iqs / 32] >> (4 * ((iqs / 16) & 1))) & 0xf;
const uint qs = data_a[a_offset + ib].qs[iqs / 8];
const float db = 0.25 * (0.5 + scale);
const uint sign7 = qs >> 9;
// Add parity bit
const uint sign8 = sign7 | (bitCount(sign7) << 7);
const uint sign = sign8 >> (iqs % 8);
const u8vec4 grid = unpack8(iq2xs_grid[qs & 511][(iqs % 8) / 4] >> (8 * (iqs % 4)));
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
return db * vec2(
grid.x * (sign0 ? -1.0 : 1.0),
grid.y * (sign1 ? -1.0 : 1.0)
);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint scale = (data_a[a_offset + ib].scales[iqs / 32] >> (4 * ((iqs / 16) & 1))) & 0xf;
const uint qs = data_a[a_offset + ib].qs[iqs / 8];
const float db = 0.25 * (0.5 + scale);
const uint sign7 = qs >> 9;
// Add parity bit
const uint sign8 = sign7 | (bitCount(sign7) << 7);
const uint sign = sign8 >> (iqs % 8);
const u8vec4 grid = unpack8(iq2xs_grid[qs & 511][(iqs % 8) / 4] >> (8 * (iqs % 4)));
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
bool sign2 = (sign & 4) != 0;
bool sign3 = (sign & 8) != 0;
return db * vec4(
grid.x * (sign0 ? -1.0 : 1.0),
grid.y * (sign1 ? -1.0 : 1.0),
grid.z * (sign2 ? -1.0 : 1.0),
grid.w * (sign3 ? -1.0 : 1.0)
);
}
#endif
#if defined(DATA_A_IQ2_S)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint ib32 = iqs / 32;
const uint ib8 = iqs / 8;
const uint scale = (data_a[a_offset + ib].scales[ib32] >> (4 * ((iqs / 16) & 1))) & 0xf;
const uint qs = data_a[a_offset + ib].qs[ib8];
const uint qh = data_a[a_offset + ib].qh[ib32];
const uint qhshift = 2 * (ib8 % 4);
const uint sign = data_a[a_offset + ib].qs[QUANT_K / 8 + ib8] >> (iqs % 8);
const float db = 0.25 * (0.5 + scale);
const u8vec4 grid = unpack8(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(iqs % 8) / 4]);
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
return db * vec2(
grid[iqs % 4] * (sign0 ? -1.0 : 1.0),
grid[(iqs % 4) + 1] * (sign1 ? -1.0 : 1.0)
);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint ib32 = iqs / 32;
const uint ib8 = iqs / 8;
const uint scale = (data_a[a_offset + ib].scales[ib32] >> (4 * ((iqs / 16) & 1))) & 0xf;
const uint qs = data_a[a_offset + ib].qs[ib8];
const uint qh = data_a[a_offset + ib].qh[ib32];
const uint qhshift = 2 * (ib8 % 4);
const uint sign = data_a[a_offset + ib].qs[QUANT_K / 8 + ib8] >> (iqs % 8);
const float db = 0.25 * (0.5 + scale);
const u8vec4 grid = unpack8(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(iqs % 8) / 4]);
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
bool sign2 = (sign & 4) != 0;
bool sign3 = (sign & 8) != 0;
return db * vec4(
grid.x * (sign0 ? -1.0 : 1.0),
grid.y * (sign1 ? -1.0 : 1.0),
grid.z * (sign2 ? -1.0 : 1.0),
grid.w * (sign3 ? -1.0 : 1.0)
);
}
#endif
#if defined(DATA_A_IQ3_XXS)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint ib4 = iqs / 4;
const uint ib32 = iqs / 32;
const uint is = QUANT_K / 4 + 4 * ib32;
const uint qs = data_a[a_offset + ib].qs[ib4];
// Scales are stored as packed 7+7+7+7+4 bits (4 sign tuples and 1 int4 scale)
const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[is / 2],
data_a_packed16[a_offset + ib].qs[is / 2 + 1]));
const float db = 0.5 * (0.5 + (signs >> 28));
const uint sign7 = bitfieldExtract(signs, 7 * (int(ib4 / 2) % 4), 7);
// Add parity bit
const uint sign8 = sign7 | (bitCount(sign7) << 7);
const uint sign = sign8 >> (iqs % 8);
const u8vec4 grid = unpack8(iq3xxs_grid[qs] >> (8 * (iqs % 4)));
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
return db * vec2(
grid.x * (sign0 ? -1.0 : 1.0),
grid.y * (sign1 ? -1.0 : 1.0)
);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint ib4 = iqs / 4;
const uint ib32 = iqs / 32;
const uint is = QUANT_K / 4 + 4 * ib32;
const uint qs = data_a[a_offset + ib].qs[ib4];
const uint signs = pack32(u16vec2(data_a_packed16[a_offset + ib].qs[is / 2],
data_a_packed16[a_offset + ib].qs[is / 2 + 1]));
const float db = 0.5 * (0.5 + (signs >> 28));
const uint sign7 = bitfieldExtract(signs, 7 * (int(ib4 / 2) % 4), 7);
// Add parity bit
const uint sign8 = sign7 | (bitCount(sign7) << 7);
const uint sign = sign8 >> (iqs % 8);
const u8vec4 grid = unpack8(iq3xxs_grid[qs]);
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
bool sign2 = (sign & 4) != 0;
bool sign3 = (sign & 8) != 0;
return db * vec4(
grid.x * (sign0 ? -1.0 : 1.0),
grid.y * (sign1 ? -1.0 : 1.0),
grid.z * (sign2 ? -1.0 : 1.0),
grid.w * (sign3 ? -1.0 : 1.0)
);
}
#endif
#if defined(DATA_A_IQ3_S)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint qs = data_a[a_offset + ib].qs[iqs / 4];
const uint qh = data_a[a_offset + ib].qh[iqs / 32];
const uint sign = data_a[a_offset + ib].signs[iqs / 8] >> (iqs % 8);
const uint scale = data_a[a_offset + ib].scales[iqs / 64];
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
const float db = 1 + 2 * ((scale >> (4 * ((iqs / 32) & 1))) & 0xf);
const uint32_t grid = iq3s_grid[qs | ((qh << (8 - ((iqs / 4) % 8))) & 256)] >> (8 * (iqs % 4));
return db * vec2(
int(grid & 0xFF) * (sign0 ? -1.0 : 1.0),
int((grid >> 8) & 0xFF) * (sign1 ? -1.0 : 1.0)
);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint ib4 = iqs / 4;
const uint ib32 = iqs / 32;
const uint qs = data_a[a_offset + ib].qs[ib4];
const uint qh = data_a[a_offset + ib].qh[ib32];
const uint sign = data_a[a_offset + ib].signs[iqs / 8] >> (iqs % 8);
const uint scale = data_a[a_offset + ib].scales[ib32 / 2];
bool sign0 = (sign & 1) != 0;
bool sign1 = (sign & 2) != 0;
bool sign2 = (sign & 4) != 0;
bool sign3 = (sign & 8) != 0;
const float db = 1 + 2 * ((scale >> (4 * (ib32 & 1))) & 0xf);
const uint32_t grid = iq3s_grid[qs | ((qh << (8 - ib4 % 8)) & 256)] >> (8 * (iqs % 4));
return db * vec4(
int(grid & 0xFF) * (sign0 ? -1.0 : 1.0),
int((grid >> 8) & 0xFF) * (sign1 ? -1.0 : 1.0),
int((grid >> 16) & 0xFF) * (sign2 ? -1.0 : 1.0),
int((grid >> 24) & 0xFF) * (sign3 ? -1.0 : 1.0)
);
}
#endif
#if defined(DATA_A_IQ4_XS)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint ib32 = iqs / 32;
const uint iq = 16 * ib32 + (iqs % 16);
const uint sl = (data_a[a_offset + ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
const uint sh = (data_a[a_offset + ib].scales_h >> (2 * ib32)) & 3;
const uint qshift = (iqs & 16) >> 2;
u8vec2 qs = u8vec2(data_a[a_offset + ib].qs[iq], data_a[a_offset + ib].qs[iq + 1]);
qs = (qs >> qshift) & uint8_t(0xF);
const float dl = float(int(sl | (sh << 4)) - 32);
return dl * vec2(kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y]);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint ib32 = iqs / 32;
const uint iq = 16 * ib32 + (iqs % 16);
const uint sl = (data_a[a_offset + ib].scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
const uint sh = (data_a[a_offset + ib].scales_h >> (2 * ib32)) & 3;
const uint qshift = (iqs & 16) >> 2;
u8vec4 qs = u8vec4(
data_a[a_offset + ib].qs[iq + 0],
data_a[a_offset + ib].qs[iq + 1],
data_a[a_offset + ib].qs[iq + 2],
data_a[a_offset + ib].qs[iq + 3]
);
qs = (qs >> qshift) & uint8_t(0xF);
const float dl = float(int(sl | (sh << 4)) - 32);
return dl * vec4(
kvalues_iq4nl[qs.x], kvalues_iq4nl[qs.y],
kvalues_iq4nl[qs.z], kvalues_iq4nl[qs.w]);
}
#endif
#if defined(DATA_A_IQ4_NL)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[vui >> 12]);
}
#endif
#if defined(DATA_A_MXFP4)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
return vec2(kvalues_mxfp4[vui & 0xF], kvalues_mxfp4[vui >> 4]);
}
vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
vec2 v0 = dequantize(ib, iqs, a_offset);
vec2 v1 = dequantize(ib, iqs + 1, a_offset);
return vec4(v0.x, v0.y, v1.x, v1.y);
}
#endif
#if defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)
vec2 get_dm(uint ib, uint a_offset) {
return vec2(0, 0);
}
#endif
#if defined(DATA_A_IQ1_M)
vec2 get_dm(uint ib, uint a_offset) {
const uint16_t[4] scales = data_a[a_offset + ib].scales;
const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
return vec2(d, 0);
}
#endif
#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ1_S) || defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
vec2 get_dm(uint ib, uint a_offset) {
return vec2(float(data_a[a_offset + ib].d), 0);
}
#endif
#if defined(DATA_A_MXFP4)
vec2 get_dm(uint ib, uint a_offset) {
return vec2(e8m0_to_fp32(data_a[a_offset + ib].e), 0);
}
#endif
#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
vec2 get_dm(uint ib, uint a_offset) {
return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
}
#endif
#if defined(DATA_A_Q2_K)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
iqs /= 2;
const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
const uint scalesi = iqs / 8; // 0..15
const uint qsshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
const uvec2 qs = uvec2(data_a[a_offset + ib].qs[qsi], data_a[a_offset + ib].qs[qsi + 1]);
const uint scales = data_a[a_offset + ib].scales[scalesi];
const vec2 d = vec2(data_a[a_offset + ib].d);
return d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
}
vec2 get_dm(uint ib, uint a_offset) {
return vec2(1, 0);
}
#endif
#if defined(DATA_A_Q3_K)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
iqs /= 2;
const uint n = iqs / 64; // 0,1
const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..62
const uint hmi = (iqs % 16) * 2; // 0,2,4..30
const uint j = (iqs % 64) / 4; // 0..3
const uint is = iqs / 8; // 0..15
const uint halfsplit = ((iqs % 64) / 16); // 0,1,2,3
const uint qsshift = halfsplit * 2; // 0,2,4,6
const uint m = 1 << (4 * n + halfsplit); // 1,2,4,8,16,32,64,128
const int8_t us = int8_t(((data_a[a_offset + ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF)
| (((data_a[a_offset + ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4));
const float dl = float(data_a[a_offset + ib].d) * float(us - 32);
return vec2(dl * float(int8_t((data_a[a_offset + ib].qs[qsi ] >> qsshift) & 3) - (((data_a[a_offset + ib].hmask[hmi ] & m) != 0) ? 0 : 4)),
dl * float(int8_t((data_a[a_offset + ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[a_offset + ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
}
vec2 get_dm(uint ib, uint a_offset) {
return vec2(1, 0);
}
#endif
#if defined(DATA_A_Q4_K)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
iqs /= 2;
const uint n = iqs / 32; // 0,1,2,3
const uint b = (iqs % 32) / 16; // 0,1
const uint is = 2 * n + b; // 0..7
const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..126
const vec2 loadd = vec2(data_a[a_offset + ib].d);
const uint scidx0 = (is < 4) ? is : (is + 4);
const uint scidx1 = (is < 4) ? is : (is - 4);
const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
const uint scidxshift1 = (is < 4) ? 0 : 2;
const uint mbidx0 = is + 4;
const uint mbidx1 = (is < 4) ? is + 4 : is;
const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
const uint mbidxshift0 = (is < 4) ? 0 : 4;
const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
const uint mbidxshift1 = (is < 4) ? 0 : 2;
const uint8_t sc = uint8_t((data_a[a_offset + ib].scales[scidx0] & 0xF) | ((data_a[a_offset + ib].scales[scidx1] & scidxmask1) >> scidxshift1));
const uint8_t mbyte = uint8_t((data_a[a_offset + ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[a_offset + ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
const float d = loadd.x * sc;
const float m = -loadd.y * mbyte;
return vec2(fma(d, float((data_a[a_offset + ib].qs[qsi ] >> (b * 4)) & 0xF), m),
fma(d, float((data_a[a_offset + ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
}
vec2 get_dm(uint ib, uint a_offset) {
return vec2(1, 0);
}
#endif
#if defined(DATA_A_Q5_K)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
iqs /= 2;
const uint n = iqs / 32; // 0,1,2,3
const uint b = (iqs % 32) / 16; // 0,1
const uint is = 2 * n + b; // 0..7
const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..126
const uint qhi = (iqs % 16) * 2; // 0,2,4..30
const uint8_t hm = uint8_t(1 << (iqs / 16));
const vec2 loadd = vec2(data_a[a_offset + ib].d);
const uint scidx0 = (is < 4) ? is : (is + 4);
const uint scidx1 = (is < 4) ? is : (is - 4);
const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
const uint scidxshift1 = (is < 4) ? 0 : 2;
const uint mbidx0 = is + 4;
const uint mbidx1 = (is < 4) ? is + 4 : is;
const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
const uint mbidxshift0 = (is < 4) ? 0 : 4;
const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
const uint mbidxshift1 = (is < 4) ? 0 : 2;
const uint8_t sc = uint8_t((data_a[a_offset + ib].scales[scidx0] & 0xF) | ((data_a[a_offset + ib].scales[scidx1] & scidxmask1) >> scidxshift1));
const uint8_t mbyte = uint8_t(((data_a[a_offset + ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((data_a[a_offset + ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
const float d = loadd.x * sc;
const float m = -loadd.y * mbyte;
return vec2(fma(d, float((data_a[a_offset + ib].qs[qsi ] >> (b * 4)) & 0xF) + float((data_a[a_offset + ib].qh[qhi ] & hm) != 0 ? 16 : 0), m),
fma(d, float((data_a[a_offset + ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[a_offset + ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
}
vec2 get_dm(uint ib, uint a_offset) {
return vec2(1, 0);
}
#endif
#if defined(DATA_A_Q6_K)
vec2 dequantize(uint ib, uint iqs, uint a_offset) {
iqs /= 2;
const uint n = iqs / 64; // 0,1
const uint b = (iqs % 64) / 32; // 0,1
const uint is_b = (iqs % 16) / 8; // 0,1
const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
const uint is = 8 * n + qhshift + is_b; // 0..15
const uint qsi = n * 64 + (iqs % 32) * 2; // 0,2,4..126
const uint qhi = n * 32 + (iqs % 16) * 2; // 0,2,4..62
const float dscale = float(data_a[a_offset + ib].d) * float(data_a[a_offset + ib].scales[is]);
return vec2(dscale * float(int8_t(((data_a[a_offset + ib].ql[qsi ] >> (b * 4)) & 0xF) | (((data_a[a_offset + ib].qh[qhi ] >> qhshift) & 3) << 4)) - 32),
dscale * float(int8_t(((data_a[a_offset + ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[a_offset + ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
}
vec2 get_dm(uint ib, uint a_offset) {
return vec2(1, 0);
}
#endif
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.glsl 0 → 100644
View file @ 2aba569a
#include "types.glsl"
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ4_0 {
block_q4_0_packed16 block;
};
float16_t dequantFuncQ4_0(const in decodeBufQ4_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint shift = (idx & 0x10) >> 2;
uint32_t qs = uint32_t(bl.block.qs[(idx & 0xE) >> 1]);
qs >>= shift;
qs &= 0x0F0F;
qs = unpack8(qs)[idx & 1];
float16_t ret = (float16_t(qs) - float16_t(8)) * d;
return ret;
}
layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ4_1 {
block_q4_1 block;
};
float16_t dequantFuncQ4_1(const in decodeBufQ4_1 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const float16_t m = bl.block.m;
const uint idx = coordInBlock[1];
const uint iqs = idx & 0xF;
const uint shift = (idx & 0x10) >> 2;
uint32_t qs = bl.block.qs[iqs];
qs >>= shift;
qs &= 0xF;
float16_t ret = float16_t(qs) * d + m;
return ret;
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ5_0 {
block_q5_0 block;
};
float16_t dequantFuncQ5_0(const in decodeBufQ5_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint iqs = idx & 0xF;
const uint uint_qh = uint(bl.block.qh[1]) << 16 | bl.block.qh[0];
const uint qh = ((uint_qh >> idx) << 4) & 0x10;
const uint shift = (idx & 0x10) >> 2;
uint32_t qs = bl.block.qs[iqs];
qs >>= shift;
qs &= 0xF;
float16_t ret = (float16_t(qs | qh) - float16_t(16)) * d;
return ret;
}
layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufQ5_1 {
block_q5_1 block;
};
float16_t dequantFuncQ5_1(const in decodeBufQ5_1 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const float16_t m = bl.block.m;
const uint idx = coordInBlock[1];
const uint iqs = idx & 0xF;
const uint uint_qh = bl.block.qh;
const uint qh = ((uint_qh >> idx) << 4) & 0x10;
const uint shift = (idx & 0x10) >> 2;
uint32_t qs = bl.block.qs[iqs];
qs >>= shift;
qs &= 0xF;
float16_t ret = float16_t(qs | qh) * d + m;
return ret;
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ8_0 {
block_q8_0_packed16 block;
};
float16_t dequantFuncQ8_0(const in decodeBufQ8_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint iqs = idx;
// Load 16b and select the byte for this element
int32_t qs = unpack8(bl.block.qs[(iqs & 0x1E) >> 1])[iqs & 1];
float16_t ret = float16_t(qs) * d;
return ret;
}
layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ2_K {
block_q2_K block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2_K_packed16 {
block_q2_K_packed16 block;
};
float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl);
const f16vec2 d = bl.block.d;
const uint idx = coordInBlock[1];
const uint scalesi = (idx & 0xF0) >> 4; // 0..15
const uint qsshift = (idx & 0x60) >> 4; // 0,2,4,6
uint qs = uint32_t(bl16.block.qs[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
qs = (qs >> qsshift) & 0x0303;
qs = unpack8(qs)[idx & 1];
const uint scales = bl.block.scales[scalesi];
float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4);
return ret;
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ3_K {
block_q3_K block;
};
float16_t dequantFuncQ3_K(const in decodeBufQ3_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const uint idx = coordInBlock[1];
const uint iqs = idx;
const uint n = iqs / 128; // 0,1
const uint qsi = n * 32 + (iqs % 32); // 0..63
const uint hmi = (iqs % 32); // 0..31
const uint j = (iqs % 128) / 8; // 0..15
const uint is = iqs / 16; // 0..15
const uint halfsplit = ((iqs % 128) / 32); // 0,1,2,3
const uint qsshift = halfsplit * 2; // 0,2,4,6
const uint m = 1 << (4 * n + halfsplit); // 1,2,4,8,16,32,64,128
uint32_t scaleidx0 = (is < 8) ? is : (is-8);
uint32_t scaleidx0shift = (is < 8) ? 0 : 4;
uint32_t scaleidx1 = is + 8 - (is/4)*4;
uint32_t scaleidx1shift = (is/4)*2;
const int8_t us = int8_t(((bl.block.scales[scaleidx0] >> scaleidx0shift) & 0xF) | (((bl.block.scales[scaleidx1] >> scaleidx1shift) & 3) << 4));
const float16_t dl = bl.block.d * float16_t(us - 32);
float16_t ret = dl * float16_t(int8_t((bl.block.qs[qsi ] >> qsshift) & 3) - (((bl.block.hmask[hmi ] & m) != 0) ? 0 : 4));
return ret;
}
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K {
block_q4_K block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed16 {
block_q4_K_packed16 block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed128 {
block_q4_K_packed128 block;
};
#if defined(IS_MUL_MM2)
// For Q4_K and Q5_K in the mat-mul shader, we decode a tile's worth of scales
// into shared memory and then process the whole tile using those scales.
// There is a fetch function that loads into private variables and then a store
// function that stores into shared memory.
// Q4_K and Q5_K have the same encoding of scales, so everything is shared except
// the part that fetches from the structure (which has a different block layout).
#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K)
const uint shAscales_stride = (BM + 2);
// 1 scale per 32 elements -> 8 scales per block, per row
shared vec2 shAscales[8 * shAscales_stride];
uvec4 row_v;
#endif
#if defined(DATA_A_Q4_K)
layout (binding = 0) readonly buffer A_Q4_K_128 {block_q4_K_packed128 data_a_q4_k_packed128[];};
void fetch_scalesQ4_K(uint ir_BM, uint pos_a, uint stride_a, uint block_k, uint tid, bool in_bounds)
{
uint tids_per_row = BLOCK_SIZE / BM;
uint is_per_tid = 8 / tids_per_row;
uint is_start = is_per_tid * (tid % tids_per_row);
uint tid_row = tid / tids_per_row;
uint row = ir_BM + tid_row;
uint block_index = pos_a + row * stride_a + (block_k / QUANT_K);
if (in_bounds || row < p.M) {
row_v = data_a_q4_k_packed128[block_index].q4k[0];
}
}
#endif
#if defined(DATA_A_Q5_K)
layout (binding = 0) readonly buffer A_Q5_K_128 {block_q5_K_packed128 data_a_q5_k_packed128[];};
void fetch_scalesQ5_K(uint ir_BM, uint pos_a, uint stride_a, uint block_k, uint tid, bool in_bounds)
{
uint tids_per_row = BLOCK_SIZE / BM;
uint is_per_tid = 8 / tids_per_row;
uint is_start = is_per_tid * (tid % tids_per_row);
uint tid_row = tid / tids_per_row;
uint row = ir_BM + tid_row;
uint block_index = pos_a + row * stride_a + (block_k / QUANT_K);
if (in_bounds || row < p.M) {
row_v = data_a_q5_k_packed128[block_index].q5k[0];
}
}
#endif
#if defined(DATA_A_Q4_K) || defined(DATA_A_Q5_K)
void store_scalesQ4_K(uint tid)
{
barrier();
uint tids_per_row = BLOCK_SIZE / BM;
uint is_per_tid = 8 / tids_per_row;
uint is_start = is_per_tid * (tid % tids_per_row);
uint tid_row = tid / tids_per_row;
[[unroll]] for (uint idx = 0; idx < is_per_tid; ++idx) {
uint is = idx + is_start;
uvec4 v = row_v;
const vec2 loadd = vec2(unpackFloat2x16(v.x));
uint32_t sc;
uint32_t mbyte;
uint32_t scale0 = v.y;
uint32_t scale4 = v.z;
uint32_t scale8 = v.w;
uint32_t sc_lo = scale0;
uint32_t mb_lo = scale4;
uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2);
uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2);
sc = is < 4 ? sc_lo : sc_hi;
mbyte = is < 4 ? mb_lo : mb_hi;
sc = sc >> (8 * (is & 3));
mbyte = mbyte >> (8 * (is & 3));
sc &= 0x3F;
mbyte &= 0x3F;
const float d = loadd.x * float(sc);
const float m = loadd.y * float(mbyte);
shAscales[is * shAscales_stride + tid_row] = vec2(d,m);
}
barrier();
}
#endif
#endif
float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ4_K_packed16 bl16 = decodeBufQ4_K_packed16(bl);
decodeBufQ4_K_packed128 bl128 = decodeBufQ4_K_packed128(bl);
const uint idx = coordInBlock[1];
const uint b = (idx & 0x20) >> 5; // 0,1
const uint is = (idx & 0xE0) >> 5; // 0..7
#if defined(IS_MUL_MM2) && defined(DATA_A_Q4_K)
vec2 v = shAscales[is * shAscales_stride + (blockCoords[0] % BM)];
float d = v.x;
float m = v.y;
#else
uvec4 v = bl128.block.q4k[0];
const vec2 loadd = vec2(unpackFloat2x16(v.x));
uint32_t sc;
uint32_t mbyte;
uint32_t scale0 = v.y;
uint32_t scale4 = v.z;
uint32_t scale8 = v.w;
uint32_t sc_lo = scale0;
uint32_t mb_lo = scale4;
uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2);
uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2);
sc = is < 4 ? sc_lo : sc_hi;
mbyte = is < 4 ? mb_lo : mb_hi;
sc = sc >> (8 * (is & 3));
mbyte = mbyte >> (8 * (is & 3));
sc &= 0x3F;
mbyte &= 0x3F;
const float d = loadd.x * float(sc);
const float m = loadd.y * float(mbyte);
#endif
uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
qs = (qs >> (b * 4 + 8 * (idx & 1))) & 0xF;
float ret = d * float(qs) - m;
return float16_t(ret);
}
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K {
block_q5_K block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed16 {
block_q5_K_packed16 block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed128 {
block_q5_K_packed128 block;
};
float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ5_K_packed16 bl16 = decodeBufQ5_K_packed16(bl);
decodeBufQ5_K_packed128 bl128 = decodeBufQ5_K_packed128(bl);
const uint idx = coordInBlock[1];
const uint b = (idx & 0x20) >> 5; // 0,1
const uint is = (idx & 0xE0) >> 5; // 0..7
#if defined(IS_MUL_MM2) && defined(DATA_A_Q5_K)
vec2 v = shAscales[is * shAscales_stride + (blockCoords[0] % BM)];
float d = v.x;
float m = v.y;
#else
uvec4 v = bl128.block.q5k[0];
const f16vec2 loadd = unpackFloat2x16(v.x);
uint32_t sc;
uint32_t mbyte;
uint32_t scale0 = v.y;
uint32_t scale4 = v.z;
uint32_t scale8 = v.w;
uint32_t sc_lo = scale0;
uint32_t mb_lo = scale4;
uint32_t sc_hi = (scale8 & 0x0F0F0F0F) | ((scale0 & 0xC0C0C0C0) >> 2);
uint32_t mb_hi = ((scale8 & 0xF0F0F0F0) >> 4) | ((scale4 & 0xC0C0C0C0) >> 2);
sc = is < 4 ? sc_lo : sc_hi;
mbyte = is < 4 ? mb_lo : mb_hi;
sc = sc >> (8 * (is & 3));
mbyte = mbyte >> (8 * (is & 3));
sc &= 0x3F;
mbyte &= 0x3F;
const float16_t d = loadd.x * float16_t(sc);
const float16_t m = loadd.y * float16_t(mbyte);
#endif
uint qh = uint32_t(bl16.block.qh[(idx & 0x1E) >> 1]);
qh = ((qh >> is) & 0x101) << 4;
uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
qs = (qs >> (b * 4)) & 0x0F0F;
qs = unpack8(qs | qh)[idx & 1];
float ret = d * float(qs) - m;
return float16_t(ret);
}
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ6_K {
block_q6_K block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ6_K_packed16 {
block_q6_K_packed16 block;
};
float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ6_K_packed16 bl16 = decodeBufQ6_K_packed16(bl);
const uint idx = coordInBlock[1];
const uint b = (idx & 0x40) >> 6; // 0,1
const uint qhshift = (idx & 0x60) >> 4; // 0,2,4,6
const uint is = (idx & 0xF0) >> 4; // 0..15
const float16_t dscale = bl.block.d * float16_t(bl.block.scales[is]);
uint ql = uint32_t(bl16.block.ql[((idx & 0x80) >> 2) + ((idx & 0x3E) >> 1)]);
ql = (ql >> (b * 4)) & 0x0F0F;
uint qh = uint32_t(bl16.block.qh[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
qh = ((qh >> qhshift) & 0x0303) << 4;
int q = unpack8(ql | qh)[idx & 1];
float16_t ret = dscale * float16_t(q - 32);
return ret;
}
#if defined(DATA_A_IQ1_S)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_S {
block_iq1_s block;
};
float16_t dequantFuncIQ1_S(const in decodeBufIQ1_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint ib32 = (idx & 0xE0) >> 5;
const uint ib8 = (idx & 0xF8) >> 3;
const uint qh = bl.block.qh[ib32];
const uint qs = bl.block.qs[ib8];
const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
const uint grid = iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)];
float16_t ret = float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * int(idx % 8), 2)) + float16_t(delta));
return ret;
}
#endif
#if defined(DATA_A_IQ1_M)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_M {
block_iq1_m block;
};
layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufIQ1_M_packed64 {
block_iq1_m_packed64 block;
};
float16_t dequantFuncIQ1_M(const in decodeBufIQ1_M bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufIQ1_M_packed64 bl64 = decodeBufIQ1_M_packed64(bl);
const uint idx = coordInBlock[1];
uvec2 scales = unpack32(bl64.block.scales);
const float16_t d = uint16BitsToHalf(uint16_t(((scales.x & 0xF000) >> 12) | ((scales.x & 0xF0000000) >> 24) | ((scales.y & 0xF000) >> 4) | ((scales.y & 0xF0000000) >> 16)));
const uint ib8 = (idx & 0xF8) >> 3;
const uint ib16 = (idx & 0xF0) >> 4;
const int i8 = int(idx % 8);
const uint sc = bl.block.scales[ib8 / 8];
const uint qs = bl.block.qs[ib8];
const uint qh = bl.block.qh[ib16] >> (4 * (ib8 & 1));
const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
const float delta = ((qh & 8) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
const uint grid = iq1s_grid[qs | ((qh & 7) << 8)];
float16_t ret = d * float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * i8, 2)) + float16_t(delta));
return ret;
}
#endif
#if defined(DATA_A_IQ2_XXS)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS {
block_iq2_xxs block;
};
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS_packed16 {
block_iq2_xxs_packed16 block;
};
float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufIQ2_XXS_packed16 bl16 = decodeBufIQ2_XXS_packed16(bl);
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint ib32 = (idx & 0xE0) >> 5; // 0..7
const uint ib8 = (idx & 0x18) >> 3; // 0..3
const uint iqs = 8 * ib32 + ib8;
const uint qs = bl.block.qs[iqs];
const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3]));
const float dscale = float(bl.block.d) * 0.25 * (0.5 + float(signscale >> 28));
uint sign = bitfieldExtract(signscale, 7 * int(ib8), 7);
sign |= bitCount(sign) << 7;
uint g2 = iq2xxs_grid[qs][(idx & 4) >> 2];
g2 >>= (idx & 2) * 8;
const vec2 g = vec2(unpack8(g2));
vec2 ret = dscale * g * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);
return float16_t(ret[idx & 1]);
}
#endif
#if defined(DATA_A_IQ2_XS)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XS {
block_iq2_xs block;
};
float16_t dequantFuncIQ2_XS(const in decodeBufIQ2_XS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint is = (idx & 0xE0) >> 5; // 0..8
const uint sshift = (idx & 0x10) >> 2; // 0,4
const uint iqs = (idx & 0xF8) >> 3; // 0..63
const uint16_t qs = bl.block.qs[iqs];
const float dscale = float(bl.block.d) * 0.25 * (0.5 + float((bl.block.scales[is] >> sshift) & 0xF));
uint sign = uint(qs >> 9);
sign |= bitCount(sign) << 7;
uint g2 = iq2xs_grid[qs & 0x1FF][(idx & 4) >> 2];
g2 >>= (idx & 2) * 8;
const vec2 g = vec2(unpack8(g2));
vec2 ret = dscale * g * ((sign & (1 << (idx & 7))) != 0 ? -1.0hf : 1.0hf);
return float16_t(ret[idx & 1]);
}
#endif
#if defined(DATA_A_IQ2_S)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_S {
block_iq2_s block;
};
float16_t dequantFuncIQ2_S(const in decodeBufIQ2_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
uint idx = coordInBlock[1];
const uint ib32 = (idx & 0xE0) >> 5; // 0..7
const uint ib8 = (idx & 0xF8) >> 3; // 0..31
const uint qhshift = 2 * (ib8 % 4);
const uint scale = (bl.block.scales[ib32] >> ((idx & 0x10) >> 2)) & 0xf;
const uint qs = bl.block.qs[ib8];
const uint qh = bl.block.qh[ib32];
const uint sign = bl.block.qs[QUANT_K / 8 + ib8] >> (idx & 0x6);
const float d = float(bl.block.d);
const float db = d * 0.25 * (0.5 + scale);
const ivec2 sign01 = 1 - (2 & ivec2(sign << 1, sign));
uint g2 = iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 4) >> 2];
g2 >>= (idx & 2) * 8;
const vec2 v = db * vec2(sign01) * vec2(unpack8(g2));
return float16_t(v[idx & 1]);
}
#endif
#if defined(DATA_A_IQ3_XXS)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_XXS {
block_iq3_xxs block;
};
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_XXS_packed16 {
block_iq3_xxs_packed16 block;
};
float16_t dequantFuncIQ3_XXS(const in decodeBufIQ3_XXS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufIQ3_XXS_packed16 bl16 = decodeBufIQ3_XXS_packed16(bl);
uint idx = coordInBlock[1];
const uint iqs = (idx & 0xFC) >> 2; // 0..63
const uint is = QUANT_K / 4 + ((idx & 0xE0) >> 3);// 8 values
const float d = float(bl.block.d);
const uint qs = bl.block.qs[iqs];
const uint signs = pack32(u16vec2(
bl16.block.qs[is/2+0],
bl16.block.qs[is/2+1]
));
const float db = d * 0.5 * (0.5 + (signs >> 28));
const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (idx & 0x6);
const ivec2 sign01 = ivec2(1 - (2 & ivec2(sign << 1, sign)));
const uint grid = iq3xxs_grid[qs] >> (16 * ((idx & 2) >> 1));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
return float16_t(v[idx & 1]);
}
#endif
#if defined(DATA_A_IQ3_S)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ3_S {
block_iq3_s block;
};
float16_t dequantFuncIQ3_S(const in decodeBufIQ3_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
uint idx = coordInBlock[1];
const uint iqs = (idx & 0xFC) >> 2; // 0..63
const uint iqh = (idx & 0xE0) >> 5;
const float d = float(bl.block.d);
const uint qs = bl.block.qs[iqs];
const uint qh = bl.block.qh[iqh];
const int8_t sign = int8_t(bl.block.signs[iqs / 2] >> (idx & 0x6));
const uint scale = bl.block.scales[iqs / 16];
const ivec2 sign01 = ivec2(1 - (2 & ivec2(sign << 1, sign)));
const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> ((idx & 2) << 3);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
return float16_t(v[idx & 1]);
}
#endif
#if defined(DATA_A_IQ4_XS)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ4_XS {
block_iq4_xs block;
};
float16_t dequantFuncIQ4_XS(const in decodeBufIQ4_XS bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint ib32 = (idx & 0xE0) >> 5; // 0..7
const uint sl = (bl.block.scales_l[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
const uint sh = ((bl.block.scales_h) >> (2 * ib32)) & 3;
const uint qshift = (idx & 16) >> 2;
const uint q = (bl.block.qs[16 * ib32 + (idx % 16)] >> qshift) & 0xF;
float16_t ret = d * float16_t(int(sl | (sh << 4)) - 32) * float16_t(kvalues_iq4nl[q]);
return ret;
}
#endif
#if defined(DATA_A_IQ4_NL)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ4_NL {
block_iq4_nl block;
};
float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint iqs = idx & 0xF;
const uint shift = (idx & 0x10) >> 2;
uint32_t qs = bl.block.qs[iqs];
qs >>= shift;
qs &= 0xF;
float16_t ret = float16_t(kvalues_iq4nl[qs]) * d;
return ret;
}
#endif
#if defined(DATA_A_MXFP4)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufMXFP4 {
block_mxfp4 block;
};
float16_t dequantFuncMXFP4(const in decodeBufMXFP4 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float d = e8m0_to_fp32(bl.block.e);
const uint idx = coordInBlock[1];
const uint iqs = idx & 0xF;
const uint shift = (idx & 0x10) >> 2;
uint32_t qs = bl.block.qs[iqs];
qs >>= shift;
qs &= 0xF;
float16_t ret = float16_t(kvalues_mxfp4[qs] * d);
return ret;
}
#endif
#if defined(DATA_A_Q4_0)
#define dequantFuncA dequantFuncQ4_0
#elif defined(DATA_A_Q4_1)
#define dequantFuncA dequantFuncQ4_1
#elif defined(DATA_A_Q5_0)
#define dequantFuncA dequantFuncQ5_0
#elif defined(DATA_A_Q5_1)
#define dequantFuncA dequantFuncQ5_1
#elif defined(DATA_A_Q8_0)
#define dequantFuncA dequantFuncQ8_0
#elif defined(DATA_A_Q2_K)
#define dequantFuncA dequantFuncQ2_K
#elif defined(DATA_A_Q3_K)
#define dequantFuncA dequantFuncQ3_K
#elif defined(DATA_A_Q4_K)
#define dequantFuncA dequantFuncQ4_K
#define fetch_scales fetch_scalesQ4_K
#define store_scales store_scalesQ4_K
#elif defined(DATA_A_Q5_K)
#define dequantFuncA dequantFuncQ5_K
#define fetch_scales fetch_scalesQ5_K
#define store_scales store_scalesQ4_K
#elif defined(DATA_A_Q6_K)
#define dequantFuncA dequantFuncQ6_K
#elif defined(DATA_A_IQ1_S)
#define dequantFuncA dequantFuncIQ1_S
#elif defined(DATA_A_IQ1_M)
#define dequantFuncA dequantFuncIQ1_M
#elif defined(DATA_A_IQ2_XXS)
#define dequantFuncA dequantFuncIQ2_XXS
#elif defined(DATA_A_IQ2_XS)
#define dequantFuncA dequantFuncIQ2_XS
#elif defined(DATA_A_IQ2_S)
#define dequantFuncA dequantFuncIQ2_S
#elif defined(DATA_A_IQ3_XXS)
#define dequantFuncA dequantFuncIQ3_XXS
#elif defined(DATA_A_IQ3_S)
#define dequantFuncA dequantFuncIQ3_S
#elif defined(DATA_A_IQ4_XS)
#define dequantFuncA dequantFuncIQ4_XS
#elif defined(DATA_A_IQ4_NL)
#define dequantFuncA dequantFuncIQ4_NL
#elif defined(DATA_A_MXFP4)
#define dequantFuncA dequantFuncMXFP4
#endif
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.glsl 0 → 100644
View file @ 2aba569a
#extension GL_EXT_control_flow_attributes : require
#extension GL_EXT_shader_16bit_storage : require
layout (push_constant) uniform parameter
{
uint M;
uint K;
uint stride_a;
uint stride_b;
uint nel;
} p;
#include "types.glsl"
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_m.comp 0 → 100644
View file @ 2aba569a
#version 450
#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
#include "dequant_head.glsl"
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_iq1_m data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
// Each thread handles 1 subblock (32 values with 2 scales)
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
init_iq_shmem(gl_WorkGroupSize);
if (ib >= p.nel / 256) {
return;
}
const uint ib32 = gl_LocalInvocationID.x % 8;
const uint ib64 = ib32 / 2;
const uint b_idx = 256 * ib + 32 * ib32;
const uint16_t[4] scales = data_a[ib].scales;
const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
const uint sc = data_a[ib].scales[ib64];
[[unroll]] for (int l = 0; l < 4; ++l) {
const uint ib16 = 2 * ib32 + l / 2;
const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1);
const uint qh = data_a[ib].qh[ib16] >> (4 * (l & 1));
const uint qs = data_a[ib].qs[4 * ib32 + l];
const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
[[unroll]] for (int j = 0; j < 8; ++j) {
data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta));
}
}
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq1_s.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "dequant_head.glsl"
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_iq1_s data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
// Each thread handles 1 subblock (32 values with 2 scales)
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
init_iq_shmem(gl_WorkGroupSize);
if (ib >= p.nel / 256) {
return;
}
const uint ib32 = gl_LocalInvocationID.x % 8;
const uint b_idx = 256 * ib + 32 * ib32;
uint qh = data_a[ib].qh[ib32];
const float d = float(data_a[ib].d);
const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
[[unroll]] for (uint l = 0; l < 4; ++l) {
const uint qs = data_a[ib].qs[4 * ib32 + l];
const uint hi = bitfieldExtract(qh, 3 * int(l), 3);
const int16_t grid = int16_t(iq1s_grid[qs | (hi << 8)]);
[[unroll]] for (int j = 0; j < 8; ++j) {
data_b[b_idx + 8 * l + j] = D_TYPE(dl * (bitfieldExtract(grid, 2*j, 2) + delta));
}
}
}
ml/backend/ggml/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp 0 → 100644
View file @ 2aba569a
#version 450
#include "dequant_head.glsl"
layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {block_iq2_s data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() {
// Each thread handles 1 subblock (32 values with 2 scales)
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
init_iq_shmem(gl_WorkGroupSize);
if (ib >= p.nel / 256) {
return;
}
const uint ib32 = gl_LocalInvocationID.x % 8;
const uint b_idx = 256 * ib + 32 * ib32;
const float d = float(data_a[ib].d);
const vec2 scale = vec2(data_a[ib].scales[ib32] & 0xf, data_a[ib].scales[ib32] >> 4);
const vec2 db = d * (0.5 + scale) * 0.25;
uint qh = data_a[ib].qh[ib32];
[[unroll]] for (uint l = 0; l < 4; ++l) {
uint qs = data_a[ib].qs[4 * ib32 + l];
const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l];
qs |= (qh << (8 - 2 * l)) & 0x300;
const uvec2 grid = iq2s_grid[qs];
const u8vec4 grid0 = unpack8(grid.x);
const u8vec4 grid1 = unpack8(grid.y);
data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 1] = D_TYPE(db[l/2] * grid0.y * ((sign & 2) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 2] = D_TYPE(db[l/2] * grid0.z * ((sign & 4) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 3] = D_TYPE(db[l/2] * grid0.w * ((sign & 8) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 4] = D_TYPE(db[l/2] * grid1.x * ((sign & 16) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 5] = D_TYPE(db[l/2] * grid1.y * ((sign & 32) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 6] = D_TYPE(db[l/2] * grid1.z * ((sign & 64) != 0 ? -1.0 : 1.0));
data_b[b_idx + 8 * l + 7] = D_TYPE(db[l/2] * grid1.w * ((sign & 128) != 0 ? -1.0 : 1.0));
}
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment