llama_cpp 0.16.2 → 0.17.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/CHANGELOG.md +18 -0
- data/README.md +7 -12
- data/ext/llama_cpp/extconf.rb +2 -43
- data/ext/llama_cpp/llama_cpp.cpp +8 -0
- data/lib/llama_cpp/version.rb +3 -3
- data/sig/llama_cpp.rbs +3 -0
- metadata +2 -171
- data/vendor/include/.gitkeep +0 -0
- data/vendor/lib/.gitkeep +0 -0
- data/vendor/tmp/llama.cpp/LICENSE +0 -21
- data/vendor/tmp/llama.cpp/Makefile +0 -1124
- data/vendor/tmp/llama.cpp/ggml-alloc.c +0 -1041
- data/vendor/tmp/llama.cpp/ggml-alloc.h +0 -76
- data/vendor/tmp/llama.cpp/ggml-backend-impl.h +0 -153
- data/vendor/tmp/llama.cpp/ggml-backend.c +0 -2225
- data/vendor/tmp/llama.cpp/ggml-backend.h +0 -236
- data/vendor/tmp/llama.cpp/ggml-blas.cpp +0 -363
- data/vendor/tmp/llama.cpp/ggml-blas.h +0 -23
- data/vendor/tmp/llama.cpp/ggml-common.h +0 -1805
- data/vendor/tmp/llama.cpp/ggml-cuda/acc.cu +0 -47
- data/vendor/tmp/llama.cpp/ggml-cuda/arange.cu +0 -34
- data/vendor/tmp/llama.cpp/ggml-cuda/argsort.cu +0 -104
- data/vendor/tmp/llama.cpp/ggml-cuda/binbcast.cu +0 -280
- data/vendor/tmp/llama.cpp/ggml-cuda/clamp.cu +0 -34
- data/vendor/tmp/llama.cpp/ggml-cuda/concat.cu +0 -196
- data/vendor/tmp/llama.cpp/ggml-cuda/convert.cu +0 -686
- data/vendor/tmp/llama.cpp/ggml-cuda/cpy.cu +0 -490
- data/vendor/tmp/llama.cpp/ggml-cuda/diagmask.cu +0 -40
- data/vendor/tmp/llama.cpp/ggml-cuda/dmmv.cu +0 -674
- data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f16.cu +0 -319
- data/vendor/tmp/llama.cpp/ggml-cuda/fattn-tile-f32.cu +0 -312
- data/vendor/tmp/llama.cpp/ggml-cuda/fattn.cu +0 -345
- data/vendor/tmp/llama.cpp/ggml-cuda/getrows.cu +0 -178
- data/vendor/tmp/llama.cpp/ggml-cuda/im2col.cu +0 -104
- data/vendor/tmp/llama.cpp/ggml-cuda/mmq.cu +0 -88
- data/vendor/tmp/llama.cpp/ggml-cuda/mmvq.cu +0 -419
- data/vendor/tmp/llama.cpp/ggml-cuda/norm.cu +0 -221
- data/vendor/tmp/llama.cpp/ggml-cuda/pad.cu +0 -49
- data/vendor/tmp/llama.cpp/ggml-cuda/pool2d.cu +0 -94
- data/vendor/tmp/llama.cpp/ggml-cuda/quantize.cu +0 -112
- data/vendor/tmp/llama.cpp/ggml-cuda/rope.cu +0 -271
- data/vendor/tmp/llama.cpp/ggml-cuda/scale.cu +0 -31
- data/vendor/tmp/llama.cpp/ggml-cuda/softmax.cu +0 -206
- data/vendor/tmp/llama.cpp/ggml-cuda/sumrows.cu +0 -40
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu +0 -10
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu +0 -9
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu +0 -10
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu +0 -10
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu +0 -8
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q2_k.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q3_k.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q4_k.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_1.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q5_k.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q6_k.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/template-instances/mmq-instance-q8_0.cu +0 -5
- data/vendor/tmp/llama.cpp/ggml-cuda/tsembd.cu +0 -47
- data/vendor/tmp/llama.cpp/ggml-cuda/unary.cu +0 -314
- data/vendor/tmp/llama.cpp/ggml-cuda/upscale.cu +0 -51
- data/vendor/tmp/llama.cpp/ggml-cuda.cu +0 -3069
- data/vendor/tmp/llama.cpp/ggml-cuda.h +0 -44
- data/vendor/tmp/llama.cpp/ggml-impl.h +0 -651
- data/vendor/tmp/llama.cpp/ggml-kompute.cpp +0 -2038
- data/vendor/tmp/llama.cpp/ggml-kompute.h +0 -46
- data/vendor/tmp/llama.cpp/ggml-metal.h +0 -66
- data/vendor/tmp/llama.cpp/ggml-metal.m +0 -3273
- data/vendor/tmp/llama.cpp/ggml-metal.metal +0 -6540
- data/vendor/tmp/llama.cpp/ggml-quants.c +0 -14994
- data/vendor/tmp/llama.cpp/ggml-quants.h +0 -133
- data/vendor/tmp/llama.cpp/ggml-rpc.cpp +0 -1178
- data/vendor/tmp/llama.cpp/ggml-rpc.h +0 -24
- data/vendor/tmp/llama.cpp/ggml-sycl.cpp +0 -6351
- data/vendor/tmp/llama.cpp/ggml-sycl.h +0 -40
- data/vendor/tmp/llama.cpp/ggml-vulkan-shaders.hpp +0 -144508
- data/vendor/tmp/llama.cpp/ggml-vulkan.cpp +0 -7183
- data/vendor/tmp/llama.cpp/ggml-vulkan.h +0 -29
- data/vendor/tmp/llama.cpp/ggml.c +0 -22506
- data/vendor/tmp/llama.cpp/ggml.h +0 -2458
- data/vendor/tmp/llama.cpp/llama.cpp +0 -18985
- data/vendor/tmp/llama.cpp/llama.h +0 -1147
- data/vendor/tmp/llama.cpp/scripts/get-flags.mk +0 -38
- data/vendor/tmp/llama.cpp/sgemm.cpp +0 -1032
- data/vendor/tmp/llama.cpp/sgemm.h +0 -14
- data/vendor/tmp/llama.cpp/unicode-data.cpp +0 -7033
- data/vendor/tmp/llama.cpp/unicode-data.h +0 -20
- data/vendor/tmp/llama.cpp/unicode.cpp +0 -810
- data/vendor/tmp/llama.cpp/unicode.h +0 -63
|
@@ -1,1032 +0,0 @@
|
|
|
1
|
-
// Copyright 2024 Mozilla Foundation
|
|
2
|
-
//
|
|
3
|
-
// Permission is hereby granted, free of charge, to any person obtaining
|
|
4
|
-
// a copy of this software and associated documentation files (the
|
|
5
|
-
// "Software"), to deal in the Software without restriction, including
|
|
6
|
-
// without limitation the rights to use, copy, modify, merge, publish,
|
|
7
|
-
// distribute, sublicense, and/or sell copies of the Software, and to
|
|
8
|
-
// permit persons to whom the Software is furnished to do so, subject to
|
|
9
|
-
// the following conditions:
|
|
10
|
-
//
|
|
11
|
-
// The above copyright notice and this permission notice shall be
|
|
12
|
-
// included in all copies or substantial portions of the Software.
|
|
13
|
-
//
|
|
14
|
-
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
|
|
15
|
-
// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
|
|
16
|
-
// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
|
|
17
|
-
// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
|
|
18
|
-
// BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
|
|
19
|
-
// ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
|
|
20
|
-
// CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
-
// SOFTWARE.
|
|
22
|
-
|
|
23
|
-
//
|
|
24
|
-
// _ _ ___ _ _ ___
|
|
25
|
-
// | |_(_)_ _ _ _| _ ) | /_\ / __|
|
|
26
|
-
// | _| | ' \ || | _ \ |__ / _ \\__ \.
|
|
27
|
-
// \__|_|_||_\_, |___/____/_/ \_\___/
|
|
28
|
-
// |__/
|
|
29
|
-
//
|
|
30
|
-
// BASIC LINEAR ALGEBRA SUBPROGRAMS
|
|
31
|
-
//
|
|
32
|
-
//
|
|
33
|
-
// This file implements multithreaded CPU matrix multiplication for the
|
|
34
|
-
// common contiguous use case C = Aᵀ * B. These kernels are designed to
|
|
35
|
-
// have excellent performance[1] for matrices that fit in the CPU cache
|
|
36
|
-
// without imposing any overhead such as cache filling or malloc calls.
|
|
37
|
-
//
|
|
38
|
-
// This implementation does not guarantee any upper bound with rounding
|
|
39
|
-
// errors, which grow along with k. Our goal's to maximally exploit the
|
|
40
|
-
// hardware for performance, and then use whatever resources remain for
|
|
41
|
-
// improving numerical accuracy.
|
|
42
|
-
//
|
|
43
|
-
// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
|
|
44
|
-
// Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
|
|
45
|
-
|
|
46
|
-
#if defined(__GNUC__)
|
|
47
|
-
#pragma GCC diagnostic ignored "-Wpedantic"
|
|
48
|
-
#pragma GCC diagnostic ignored "-Wignored-attributes"
|
|
49
|
-
#endif
|
|
50
|
-
|
|
51
|
-
#include "sgemm.h"
|
|
52
|
-
#include "ggml-impl.h"
|
|
53
|
-
#include "ggml-quants.h"
|
|
54
|
-
|
|
55
|
-
#ifdef _MSC_VER
|
|
56
|
-
#define NOINLINE __declspec(noinline)
|
|
57
|
-
#else
|
|
58
|
-
#define NOINLINE __attribute__((__noinline__))
|
|
59
|
-
#endif
|
|
60
|
-
|
|
61
|
-
#if defined(__ARM_NEON) || defined(__AVX512F__)
|
|
62
|
-
#define VECTOR_REGISTERS 32
|
|
63
|
-
#else
|
|
64
|
-
#define VECTOR_REGISTERS 16
|
|
65
|
-
#endif
|
|
66
|
-
|
|
67
|
-
#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
|
|
68
|
-
|
|
69
|
-
namespace {
|
|
70
|
-
|
|
71
|
-
inline float unhalf(ggml_fp16_t d) {
|
|
72
|
-
return GGML_FP16_TO_FP32(d);
|
|
73
|
-
}
|
|
74
|
-
|
|
75
|
-
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
76
|
-
// VECTORIZED ARITHMETIC OPERATIONS
|
|
77
|
-
|
|
78
|
-
#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
|
79
|
-
inline __m128 add(__m128 x, __m128 y) { return _mm_add_ps(x, y); }
|
|
80
|
-
inline __m128 sub(__m128 x, __m128 y) { return _mm_sub_ps(x, y); }
|
|
81
|
-
inline __m128 mul(__m128 x, __m128 y) { return _mm_mul_ps(x, y); }
|
|
82
|
-
#endif // __SSE__
|
|
83
|
-
|
|
84
|
-
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
|
85
|
-
inline __m256 add(__m256 x, __m256 y) { return _mm256_add_ps(x, y); }
|
|
86
|
-
inline __m256 sub(__m256 x, __m256 y) { return _mm256_sub_ps(x, y); }
|
|
87
|
-
inline __m256 mul(__m256 x, __m256 y) { return _mm256_mul_ps(x, y); }
|
|
88
|
-
#endif // __AVX__
|
|
89
|
-
|
|
90
|
-
#if defined(__AVX512F__)
|
|
91
|
-
inline __m512 add(__m512 x, __m512 y) { return _mm512_add_ps(x, y); }
|
|
92
|
-
inline __m512 sub(__m512 x, __m512 y) { return _mm512_sub_ps(x, y); }
|
|
93
|
-
inline __m512 mul(__m512 x, __m512 y) { return _mm512_mul_ps(x, y); }
|
|
94
|
-
#endif // __AVX512F__
|
|
95
|
-
|
|
96
|
-
#if defined(__ARM_NEON)
|
|
97
|
-
inline float32x4_t add(float32x4_t x, float32x4_t y) { return vaddq_f32(x, y); }
|
|
98
|
-
inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vsubq_f32(x, y); }
|
|
99
|
-
inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vmulq_f32(x, y); }
|
|
100
|
-
#endif // __ARM_NEON
|
|
101
|
-
|
|
102
|
-
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
|
|
103
|
-
inline float16x8_t add(float16x8_t x, float16x8_t y) { return vaddq_f16(x, y); }
|
|
104
|
-
inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
|
|
105
|
-
inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
|
|
106
|
-
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
|
|
107
|
-
|
|
108
|
-
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
109
|
-
// VECTORIZED FUSED MULTIPLY ADD
|
|
110
|
-
|
|
111
|
-
/**
|
|
112
|
-
* Computes a * b + c.
|
|
113
|
-
*/
|
|
114
|
-
template <typename T, typename U>
|
|
115
|
-
inline U madd(T a, T b, U c) {
|
|
116
|
-
return add(mul(a, b), c);
|
|
117
|
-
}
|
|
118
|
-
|
|
119
|
-
#if defined(__FMA__)
|
|
120
|
-
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
|
121
|
-
template <>
|
|
122
|
-
inline __m256 madd(__m256 a, __m256 b, __m256 c) {
|
|
123
|
-
return _mm256_fmadd_ps(a, b, c);
|
|
124
|
-
}
|
|
125
|
-
#endif
|
|
126
|
-
#if defined(__AVX512F__)
|
|
127
|
-
template <>
|
|
128
|
-
inline __m512 madd(__m512 a, __m512 b, __m512 c) {
|
|
129
|
-
return _mm512_fmadd_ps(a, b, c);
|
|
130
|
-
}
|
|
131
|
-
#endif
|
|
132
|
-
#endif
|
|
133
|
-
|
|
134
|
-
#if defined(__ARM_FEATURE_FMA)
|
|
135
|
-
template <>
|
|
136
|
-
inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
|
|
137
|
-
return vfmaq_f32(c, b, a);
|
|
138
|
-
}
|
|
139
|
-
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
|
|
140
|
-
template <>
|
|
141
|
-
inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
|
|
142
|
-
return vfmaq_f16(c, b, a);
|
|
143
|
-
}
|
|
144
|
-
#endif
|
|
145
|
-
#endif
|
|
146
|
-
|
|
147
|
-
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
148
|
-
// VECTORIZED HORIZONTAL SUM
|
|
149
|
-
|
|
150
|
-
#if defined(__ARM_NEON)
|
|
151
|
-
inline float hsum(float32x4_t x) {
|
|
152
|
-
return vaddvq_f32(x);
|
|
153
|
-
}
|
|
154
|
-
#endif // __ARM_NEON
|
|
155
|
-
|
|
156
|
-
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
|
|
157
|
-
inline float hsum(float16x8_t x) {
|
|
158
|
-
return vaddvq_f32(vaddq_f32(vcvt_f32_f16(vget_low_f16(x)),
|
|
159
|
-
vcvt_f32_f16(vget_high_f16(x))));
|
|
160
|
-
}
|
|
161
|
-
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
|
|
162
|
-
|
|
163
|
-
#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
|
164
|
-
inline float hsum(__m128 x) {
|
|
165
|
-
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
|
166
|
-
x = _mm_add_ps(x, _mm_movehl_ps(x, x));
|
|
167
|
-
x = _mm_add_ss(x, _mm_movehdup_ps(x));
|
|
168
|
-
#else
|
|
169
|
-
__m128 t;
|
|
170
|
-
t = _mm_shuffle_ps(x, x, _MM_SHUFFLE(2, 3, 0, 1));
|
|
171
|
-
x = _mm_add_ps(x, t);
|
|
172
|
-
t = _mm_movehl_ps(t, x);
|
|
173
|
-
x = _mm_add_ss(x, t);
|
|
174
|
-
#endif
|
|
175
|
-
return _mm_cvtss_f32(x);
|
|
176
|
-
}
|
|
177
|
-
#endif
|
|
178
|
-
|
|
179
|
-
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
|
180
|
-
inline float hsum(__m256 x) {
|
|
181
|
-
return hsum(_mm_add_ps(_mm256_extractf128_ps(x, 1),
|
|
182
|
-
_mm256_castps256_ps128(x)));
|
|
183
|
-
}
|
|
184
|
-
#endif // __AVX__
|
|
185
|
-
|
|
186
|
-
#if defined(__AVX512F__)
|
|
187
|
-
inline float hsum(__m512 x) {
|
|
188
|
-
return _mm512_reduce_add_ps(x);
|
|
189
|
-
}
|
|
190
|
-
#endif // __AVX512F__
|
|
191
|
-
|
|
192
|
-
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
193
|
-
// VECTORIZED MEMORY LOADING
|
|
194
|
-
|
|
195
|
-
template <typename T, typename U> T load(const U *);
|
|
196
|
-
|
|
197
|
-
#if defined(__ARM_NEON)
|
|
198
|
-
template <> inline float32x4_t load(const float *p) {
|
|
199
|
-
return vld1q_f32(p);
|
|
200
|
-
}
|
|
201
|
-
#if !defined(_MSC_VER)
|
|
202
|
-
template <> inline float16x8_t load(const ggml_fp16_t *p) {
|
|
203
|
-
return vld1q_f16((const float16_t *)p);
|
|
204
|
-
}
|
|
205
|
-
template <> inline float32x4_t load(const ggml_fp16_t *p) {
|
|
206
|
-
return vcvt_f32_f16(vld1_f16((const float16_t *)p));
|
|
207
|
-
}
|
|
208
|
-
#endif // _MSC_VER
|
|
209
|
-
#endif // __ARM_NEON
|
|
210
|
-
|
|
211
|
-
#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
|
212
|
-
template <> inline __m128 load(const float *p) {
|
|
213
|
-
return _mm_loadu_ps(p);
|
|
214
|
-
}
|
|
215
|
-
#endif // __SSE__
|
|
216
|
-
|
|
217
|
-
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
|
|
218
|
-
template <> inline __m256 load(const float *p) {
|
|
219
|
-
return _mm256_loadu_ps(p);
|
|
220
|
-
}
|
|
221
|
-
#endif // __AVX__
|
|
222
|
-
|
|
223
|
-
#if defined(__F16C__)
|
|
224
|
-
template <> inline __m256 load(const ggml_fp16_t *p) {
|
|
225
|
-
return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
|
|
226
|
-
}
|
|
227
|
-
#endif // __F16C__
|
|
228
|
-
|
|
229
|
-
#if defined(__AVX512F__)
|
|
230
|
-
template <> inline __m512 load(const float *p) {
|
|
231
|
-
return _mm512_loadu_ps(p);
|
|
232
|
-
}
|
|
233
|
-
template <> inline __m512 load(const ggml_fp16_t *p) {
|
|
234
|
-
return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
|
|
235
|
-
}
|
|
236
|
-
#endif // __AVX512F__
|
|
237
|
-
|
|
238
|
-
////////////////////////////////////////////////////////////////////////////////////////////////////
|
|
239
|
-
// FLOATING POINT MATRIX MULTIPLICATION
|
|
240
|
-
|
|
241
|
-
template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
|
|
242
|
-
class tinyBLAS {
|
|
243
|
-
public:
|
|
244
|
-
tinyBLAS(int64_t k,
|
|
245
|
-
const TA *A, int64_t lda,
|
|
246
|
-
const TB *B, int64_t ldb,
|
|
247
|
-
TC *C, int64_t ldc,
|
|
248
|
-
int ith, int nth)
|
|
249
|
-
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
|
|
250
|
-
}
|
|
251
|
-
|
|
252
|
-
void matmul(int64_t m, int64_t n, int task) {
|
|
253
|
-
if (task == GGML_TASK_TYPE_COMPUTE)
|
|
254
|
-
mnpack(0, m, 0, n);
|
|
255
|
-
}
|
|
256
|
-
|
|
257
|
-
private:
|
|
258
|
-
NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
|
|
259
|
-
int64_t mc, nc, mp, np;
|
|
260
|
-
switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
|
|
261
|
-
#if VECTOR_REGISTERS == 32
|
|
262
|
-
case 0x55:
|
|
263
|
-
mc = 5;
|
|
264
|
-
nc = 5;
|
|
265
|
-
gemm<5, 5>(m0, m, n0, n);
|
|
266
|
-
break;
|
|
267
|
-
case 0x45:
|
|
268
|
-
mc = 4;
|
|
269
|
-
nc = 5;
|
|
270
|
-
gemm<4, 5>(m0, m, n0, n);
|
|
271
|
-
break;
|
|
272
|
-
case 0x54:
|
|
273
|
-
mc = 5;
|
|
274
|
-
nc = 4;
|
|
275
|
-
gemm<5, 4>(m0, m, n0, n);
|
|
276
|
-
break;
|
|
277
|
-
case 0x44:
|
|
278
|
-
mc = 4;
|
|
279
|
-
nc = 4;
|
|
280
|
-
gemm<4, 4>(m0, m, n0, n);
|
|
281
|
-
break;
|
|
282
|
-
case 0x53:
|
|
283
|
-
mc = 5;
|
|
284
|
-
nc = 3;
|
|
285
|
-
gemm<5, 3>(m0, m, n0, n);
|
|
286
|
-
break;
|
|
287
|
-
case 0x35:
|
|
288
|
-
mc = 3;
|
|
289
|
-
nc = 5;
|
|
290
|
-
gemm<3, 5>(m0, m, n0, n);
|
|
291
|
-
break;
|
|
292
|
-
case 0x43:
|
|
293
|
-
mc = 4;
|
|
294
|
-
nc = 3;
|
|
295
|
-
gemm<4, 3>(m0, m, n0, n);
|
|
296
|
-
break;
|
|
297
|
-
#else
|
|
298
|
-
case 0x55:
|
|
299
|
-
case 0x54:
|
|
300
|
-
case 0x53:
|
|
301
|
-
case 0x45:
|
|
302
|
-
case 0x44:
|
|
303
|
-
case 0x43:
|
|
304
|
-
mc = 4;
|
|
305
|
-
nc = 3;
|
|
306
|
-
gemm<4, 3>(m0, m, n0, n);
|
|
307
|
-
break;
|
|
308
|
-
case 0x35:
|
|
309
|
-
#endif
|
|
310
|
-
case 0x34:
|
|
311
|
-
mc = 3;
|
|
312
|
-
nc = 4;
|
|
313
|
-
gemm<3, 4>(m0, m, n0, n);
|
|
314
|
-
break;
|
|
315
|
-
case 0x52:
|
|
316
|
-
mc = 5;
|
|
317
|
-
nc = 2;
|
|
318
|
-
gemm<5, 2>(m0, m, n0, n);
|
|
319
|
-
break;
|
|
320
|
-
case 0x33:
|
|
321
|
-
mc = 3;
|
|
322
|
-
nc = 3;
|
|
323
|
-
gemm<3, 3>(m0, m, n0, n);
|
|
324
|
-
break;
|
|
325
|
-
case 0x25:
|
|
326
|
-
mc = 2;
|
|
327
|
-
nc = 5;
|
|
328
|
-
gemm<2, 5>(m0, m, n0, n);
|
|
329
|
-
break;
|
|
330
|
-
case 0x42:
|
|
331
|
-
mc = 4;
|
|
332
|
-
nc = 2;
|
|
333
|
-
gemm<4, 2>(m0, m, n0, n);
|
|
334
|
-
break;
|
|
335
|
-
case 0x24:
|
|
336
|
-
mc = 2;
|
|
337
|
-
nc = 4;
|
|
338
|
-
gemm<2, 4>(m0, m, n0, n);
|
|
339
|
-
break;
|
|
340
|
-
case 0x32:
|
|
341
|
-
mc = 3;
|
|
342
|
-
nc = 2;
|
|
343
|
-
gemm<3, 2>(m0, m, n0, n);
|
|
344
|
-
break;
|
|
345
|
-
case 0x23:
|
|
346
|
-
mc = 2;
|
|
347
|
-
nc = 3;
|
|
348
|
-
gemm<2, 3>(m0, m, n0, n);
|
|
349
|
-
break;
|
|
350
|
-
case 0x51:
|
|
351
|
-
mc = 5;
|
|
352
|
-
nc = 1;
|
|
353
|
-
gemm<5, 1>(m0, m, n0, n);
|
|
354
|
-
break;
|
|
355
|
-
case 0x41:
|
|
356
|
-
mc = 4;
|
|
357
|
-
nc = 1;
|
|
358
|
-
gemm<4, 1>(m0, m, n0, n);
|
|
359
|
-
break;
|
|
360
|
-
case 0x22:
|
|
361
|
-
mc = 2;
|
|
362
|
-
nc = 2;
|
|
363
|
-
gemm<2, 2>(m0, m, n0, n);
|
|
364
|
-
break;
|
|
365
|
-
case 0x15:
|
|
366
|
-
mc = 1;
|
|
367
|
-
nc = 5;
|
|
368
|
-
gemm<1, 5>(m0, m, n0, n);
|
|
369
|
-
break;
|
|
370
|
-
case 0x14:
|
|
371
|
-
mc = 1;
|
|
372
|
-
nc = 4;
|
|
373
|
-
gemm<1, 4>(m0, m, n0, n);
|
|
374
|
-
break;
|
|
375
|
-
case 0x31:
|
|
376
|
-
mc = 3;
|
|
377
|
-
nc = 1;
|
|
378
|
-
gemm<3, 1>(m0, m, n0, n);
|
|
379
|
-
break;
|
|
380
|
-
case 0x13:
|
|
381
|
-
mc = 1;
|
|
382
|
-
nc = 3;
|
|
383
|
-
gemm<1, 3>(m0, m, n0, n);
|
|
384
|
-
break;
|
|
385
|
-
case 0x21:
|
|
386
|
-
mc = 2;
|
|
387
|
-
nc = 1;
|
|
388
|
-
gemm<2, 1>(m0, m, n0, n);
|
|
389
|
-
break;
|
|
390
|
-
case 0x12:
|
|
391
|
-
mc = 1;
|
|
392
|
-
nc = 2;
|
|
393
|
-
gemm<1, 2>(m0, m, n0, n);
|
|
394
|
-
break;
|
|
395
|
-
case 0x11:
|
|
396
|
-
mc = 1;
|
|
397
|
-
nc = 1;
|
|
398
|
-
gemm<1, 1>(m0, m, n0, n);
|
|
399
|
-
break;
|
|
400
|
-
default:
|
|
401
|
-
return;
|
|
402
|
-
}
|
|
403
|
-
mp = m0 + (m - m0) / mc * mc;
|
|
404
|
-
np = n0 + (n - n0) / nc * nc;
|
|
405
|
-
mnpack(mp, m, n0, np);
|
|
406
|
-
mnpack(m0, m, np, n);
|
|
407
|
-
}
|
|
408
|
-
|
|
409
|
-
template <int RM, int RN>
|
|
410
|
-
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
|
|
411
|
-
int64_t ytiles = (m - m0) / RM;
|
|
412
|
-
int64_t xtiles = (n - n0) / RN;
|
|
413
|
-
int64_t tiles = xtiles * ytiles;
|
|
414
|
-
int64_t duty = (tiles + nth - 1) / nth;
|
|
415
|
-
int64_t start = duty * ith;
|
|
416
|
-
int64_t end = start + duty;
|
|
417
|
-
if (end > tiles)
|
|
418
|
-
end = tiles;
|
|
419
|
-
for (int64_t job = start; job < end; ++job) {
|
|
420
|
-
int64_t ii = m0 + job / xtiles * RM;
|
|
421
|
-
int64_t jj = n0 + job % xtiles * RN;
|
|
422
|
-
D Cv[RN][RM] = {};
|
|
423
|
-
for (int64_t l = 0; l < k; l += KN)
|
|
424
|
-
for (int64_t j = 0; j < RN; ++j)
|
|
425
|
-
for (int64_t i = 0; i < RM; ++i)
|
|
426
|
-
Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
|
|
427
|
-
load<V>(B + ldb * (jj + j) + l),
|
|
428
|
-
Cv[j][i]);
|
|
429
|
-
for (int64_t j = 0; j < RN; ++j)
|
|
430
|
-
for (int64_t i = 0; i < RM; ++i)
|
|
431
|
-
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
|
|
432
|
-
}
|
|
433
|
-
}
|
|
434
|
-
|
|
435
|
-
const TA *const A;
|
|
436
|
-
const TB *const B;
|
|
437
|
-
TC *const C;
|
|
438
|
-
const int64_t k;
|
|
439
|
-
const int64_t lda;
|
|
440
|
-
const int64_t ldb;
|
|
441
|
-
const int64_t ldc;
|
|
442
|
-
const int ith;
|
|
443
|
-
const int nth;
|
|
444
|
-
};
|
|
445
|
-
|
|
446
|
-
//////////////////////////////////////////////////////////////////////////////////////////
|
|
447
|
-
// QUANT ZERO MATRIX MULTIPLICATION
|
|
448
|
-
|
|
449
|
-
#if defined(__ARM_FEATURE_DOTPROD)
|
|
450
|
-
template <typename TA>
|
|
451
|
-
class tinyBLAS_Q0_ARM {
|
|
452
|
-
public:
|
|
453
|
-
tinyBLAS_Q0_ARM(int64_t k,
|
|
454
|
-
const TA *A, int64_t lda,
|
|
455
|
-
const block_q8_0 *B, int64_t ldb,
|
|
456
|
-
float *C, int64_t ldc,
|
|
457
|
-
int ith, int nth)
|
|
458
|
-
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
|
|
459
|
-
}
|
|
460
|
-
|
|
461
|
-
void matmul(int64_t m, int64_t n, int task) {
|
|
462
|
-
if (task == GGML_TASK_TYPE_COMPUTE)
|
|
463
|
-
mnpack(0, m, 0, n);
|
|
464
|
-
}
|
|
465
|
-
|
|
466
|
-
private:
|
|
467
|
-
NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
|
|
468
|
-
int64_t mc, nc, mp, np;
|
|
469
|
-
switch ((MIN(m - m0, 3) << 4) | MIN(n - n0, 3ll)) {
|
|
470
|
-
case 0x33:
|
|
471
|
-
mc = 3;
|
|
472
|
-
nc = 3;
|
|
473
|
-
gemm<3, 3>(m0, m, n0, n);
|
|
474
|
-
break;
|
|
475
|
-
case 0x32:
|
|
476
|
-
mc = 3;
|
|
477
|
-
nc = 2;
|
|
478
|
-
gemm<3, 2>(m0, m, n0, n);
|
|
479
|
-
break;
|
|
480
|
-
case 0x23:
|
|
481
|
-
mc = 2;
|
|
482
|
-
nc = 3;
|
|
483
|
-
gemm<2, 3>(m0, m, n0, n);
|
|
484
|
-
break;
|
|
485
|
-
case 0x22:
|
|
486
|
-
mc = 2;
|
|
487
|
-
nc = 2;
|
|
488
|
-
gemm<2, 2>(m0, m, n0, n);
|
|
489
|
-
break;
|
|
490
|
-
case 0x31:
|
|
491
|
-
mc = 3;
|
|
492
|
-
nc = 1;
|
|
493
|
-
gemm<3, 1>(m0, m, n0, n);
|
|
494
|
-
break;
|
|
495
|
-
case 0x13:
|
|
496
|
-
mc = 1;
|
|
497
|
-
nc = 3;
|
|
498
|
-
gemm<1, 3>(m0, m, n0, n);
|
|
499
|
-
break;
|
|
500
|
-
case 0x21:
|
|
501
|
-
mc = 2;
|
|
502
|
-
nc = 1;
|
|
503
|
-
gemm<2, 1>(m0, m, n0, n);
|
|
504
|
-
break;
|
|
505
|
-
case 0x12:
|
|
506
|
-
mc = 1;
|
|
507
|
-
nc = 2;
|
|
508
|
-
gemm<1, 2>(m0, m, n0, n);
|
|
509
|
-
break;
|
|
510
|
-
case 0x11:
|
|
511
|
-
mc = 1;
|
|
512
|
-
nc = 1;
|
|
513
|
-
gemm<1, 1>(m0, m, n0, n);
|
|
514
|
-
break;
|
|
515
|
-
default:
|
|
516
|
-
return;
|
|
517
|
-
}
|
|
518
|
-
mp = m0 + (m - m0) / mc * mc;
|
|
519
|
-
np = n0 + (n - n0) / nc * nc;
|
|
520
|
-
mnpack(mp, m, n0, np);
|
|
521
|
-
mnpack(m0, m, np, n);
|
|
522
|
-
}
|
|
523
|
-
|
|
524
|
-
template <int RM, int RN>
|
|
525
|
-
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
|
|
526
|
-
int64_t ytiles = (m - m0) / RM;
|
|
527
|
-
int64_t xtiles = (n - n0) / RN;
|
|
528
|
-
int64_t tiles = xtiles * ytiles;
|
|
529
|
-
int64_t duty = (tiles + nth - 1) / nth;
|
|
530
|
-
int64_t start = duty * ith;
|
|
531
|
-
int64_t end = start + duty;
|
|
532
|
-
if (end > tiles)
|
|
533
|
-
end = tiles;
|
|
534
|
-
for (int64_t job = start; job < end; ++job) {
|
|
535
|
-
int64_t ii = m0 + job / xtiles * RM;
|
|
536
|
-
int64_t jj = n0 + job % xtiles * RN;
|
|
537
|
-
float32x4_t Cv[RN][RM] = {};
|
|
538
|
-
for (int64_t l = 0; l < k; ++l)
|
|
539
|
-
for (int64_t j = 0; j < RN; ++j)
|
|
540
|
-
for (int64_t i = 0; i < RM; ++i)
|
|
541
|
-
Cv[j][i] = vmlaq_n_f32(Cv[j][i],
|
|
542
|
-
vcvtq_f32_s32(vdotq_s32(
|
|
543
|
-
vdotq_s32(vdupq_n_s32(0),
|
|
544
|
-
load_lo(A + lda * (ii + i) + l),
|
|
545
|
-
load_lo(B + ldb * (jj + j) + l)),
|
|
546
|
-
load_hi(A + lda * (ii + i) + l),
|
|
547
|
-
load_hi(B + ldb * (jj + j) + l))),
|
|
548
|
-
unhalf(A[lda * (ii + i) + l].d) *
|
|
549
|
-
unhalf(B[ldb * (jj + j) + l].d));
|
|
550
|
-
for (int64_t j = 0; j < RN; ++j)
|
|
551
|
-
for (int64_t i = 0; i < RM; ++i)
|
|
552
|
-
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
|
|
553
|
-
}
|
|
554
|
-
}
|
|
555
|
-
|
|
556
|
-
inline int8x16_t load_lo(const block_q8_0 *b) {
|
|
557
|
-
return vld1q_s8(b->qs);
|
|
558
|
-
}
|
|
559
|
-
|
|
560
|
-
inline int8x16_t load_hi(const block_q8_0 *b) {
|
|
561
|
-
return vld1q_s8(b->qs + 16);
|
|
562
|
-
}
|
|
563
|
-
|
|
564
|
-
inline int8x16_t load_lo(const block_q4_0 *b) {
|
|
565
|
-
return vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vld1q_u8(b->qs),
|
|
566
|
-
vdupq_n_u8(0x0f))),
|
|
567
|
-
vdupq_n_s8(0x8));
|
|
568
|
-
}
|
|
569
|
-
|
|
570
|
-
inline int8x16_t load_hi(const block_q4_0 *b) {
|
|
571
|
-
return vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(vld1q_u8(b->qs), 4)),
|
|
572
|
-
vdupq_n_s8(0x8));
|
|
573
|
-
}
|
|
574
|
-
|
|
575
|
-
const TA *const A;
|
|
576
|
-
const block_q8_0 *const B;
|
|
577
|
-
float *const C;
|
|
578
|
-
const int64_t k;
|
|
579
|
-
const int64_t lda;
|
|
580
|
-
const int64_t ldb;
|
|
581
|
-
const int64_t ldc;
|
|
582
|
-
const int ith;
|
|
583
|
-
const int nth;
|
|
584
|
-
};
|
|
585
|
-
#endif // __ARM_FEATURE_DOTPROD
|
|
586
|
-
|
|
587
|
-
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
|
|
588
|
-
template <typename TA, typename TB, typename TC>
|
|
589
|
-
class tinyBLAS_Q0_AVX {
|
|
590
|
-
public:
|
|
591
|
-
tinyBLAS_Q0_AVX(int64_t k,
|
|
592
|
-
const TA *A, int64_t lda,
|
|
593
|
-
const TB *B, int64_t ldb,
|
|
594
|
-
TC *C, int64_t ldc,
|
|
595
|
-
int ith, int nth)
|
|
596
|
-
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
|
|
597
|
-
}
|
|
598
|
-
|
|
599
|
-
void matmul(int64_t m, int64_t n, int task) {
|
|
600
|
-
if (task == GGML_TASK_TYPE_COMPUTE)
|
|
601
|
-
mnpack(0, m, 0, n);
|
|
602
|
-
}
|
|
603
|
-
|
|
604
|
-
private:
|
|
605
|
-
void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
|
|
606
|
-
int64_t mc, nc, mp, np;
|
|
607
|
-
switch ((MIN(m - m0, 4) << 4) | MIN(n - n0, 4)) {
|
|
608
|
-
#if VECTOR_REGISTERS == 32
|
|
609
|
-
case 0x44:
|
|
610
|
-
mc = 4;
|
|
611
|
-
nc = 4;
|
|
612
|
-
gemm<4, 4>(m0, m, n0, n);
|
|
613
|
-
break;
|
|
614
|
-
case 0x43:
|
|
615
|
-
mc = 4;
|
|
616
|
-
nc = 3;
|
|
617
|
-
gemm<4, 3>(m0, m, n0, n);
|
|
618
|
-
break;
|
|
619
|
-
case 0x34:
|
|
620
|
-
mc = 3;
|
|
621
|
-
nc = 4;
|
|
622
|
-
gemm<3, 4>(m0, m, n0, n);
|
|
623
|
-
break;
|
|
624
|
-
case 0x33:
|
|
625
|
-
mc = 3;
|
|
626
|
-
nc = 3;
|
|
627
|
-
gemm<3, 3>(m0, m, n0, n);
|
|
628
|
-
break;
|
|
629
|
-
case 0x42:
|
|
630
|
-
mc = 4;
|
|
631
|
-
nc = 2;
|
|
632
|
-
gemm<4, 2>(m0, m, n0, n);
|
|
633
|
-
break;
|
|
634
|
-
case 0x24:
|
|
635
|
-
mc = 2;
|
|
636
|
-
nc = 4;
|
|
637
|
-
gemm<2, 4>(m0, m, n0, n);
|
|
638
|
-
break;
|
|
639
|
-
#else
|
|
640
|
-
case 0x44:
|
|
641
|
-
case 0x43:
|
|
642
|
-
case 0x42:
|
|
643
|
-
mc = 4;
|
|
644
|
-
nc = 2;
|
|
645
|
-
gemm<4, 2>(m0, m, n0, n);
|
|
646
|
-
break;
|
|
647
|
-
case 0x34:
|
|
648
|
-
case 0x24:
|
|
649
|
-
mc = 2;
|
|
650
|
-
nc = 4;
|
|
651
|
-
gemm<2, 4>(m0, m, n0, n);
|
|
652
|
-
break;
|
|
653
|
-
case 0x33:
|
|
654
|
-
#endif
|
|
655
|
-
case 0x32:
|
|
656
|
-
mc = 3;
|
|
657
|
-
nc = 2;
|
|
658
|
-
gemm<3, 2>(m0, m, n0, n);
|
|
659
|
-
break;
|
|
660
|
-
case 0x23:
|
|
661
|
-
mc = 2;
|
|
662
|
-
nc = 3;
|
|
663
|
-
gemm<2, 3>(m0, m, n0, n);
|
|
664
|
-
break;
|
|
665
|
-
case 0x41:
|
|
666
|
-
mc = 4;
|
|
667
|
-
nc = 1;
|
|
668
|
-
gemm<4, 1>(m0, m, n0, n);
|
|
669
|
-
break;
|
|
670
|
-
case 0x22:
|
|
671
|
-
mc = 2;
|
|
672
|
-
nc = 2;
|
|
673
|
-
gemm<2, 2>(m0, m, n0, n);
|
|
674
|
-
break;
|
|
675
|
-
case 0x14:
|
|
676
|
-
mc = 1;
|
|
677
|
-
nc = 4;
|
|
678
|
-
gemm<1, 4>(m0, m, n0, n);
|
|
679
|
-
break;
|
|
680
|
-
case 0x31:
|
|
681
|
-
mc = 3;
|
|
682
|
-
nc = 1;
|
|
683
|
-
gemm<3, 1>(m0, m, n0, n);
|
|
684
|
-
break;
|
|
685
|
-
case 0x13:
|
|
686
|
-
mc = 1;
|
|
687
|
-
nc = 3;
|
|
688
|
-
gemm<1, 3>(m0, m, n0, n);
|
|
689
|
-
break;
|
|
690
|
-
case 0x21:
|
|
691
|
-
mc = 2;
|
|
692
|
-
nc = 1;
|
|
693
|
-
gemm<2, 1>(m0, m, n0, n);
|
|
694
|
-
break;
|
|
695
|
-
case 0x12:
|
|
696
|
-
mc = 1;
|
|
697
|
-
nc = 2;
|
|
698
|
-
gemm<1, 2>(m0, m, n0, n);
|
|
699
|
-
break;
|
|
700
|
-
case 0x11:
|
|
701
|
-
mc = 1;
|
|
702
|
-
nc = 1;
|
|
703
|
-
gemm<1, 1>(m0, m, n0, n);
|
|
704
|
-
break;
|
|
705
|
-
default:
|
|
706
|
-
return;
|
|
707
|
-
}
|
|
708
|
-
mp = m0 + (m - m0) / mc * mc;
|
|
709
|
-
np = n0 + (n - n0) / nc * nc;
|
|
710
|
-
mnpack(mp, m, n0, np);
|
|
711
|
-
mnpack(m0, m, np, n);
|
|
712
|
-
}
|
|
713
|
-
|
|
714
|
-
template <int RM, int RN>
|
|
715
|
-
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
|
|
716
|
-
int64_t ytiles = (m - m0) / RM;
|
|
717
|
-
int64_t xtiles = (n - n0) / RN;
|
|
718
|
-
int64_t tiles = xtiles * ytiles;
|
|
719
|
-
int64_t duty = (tiles + nth - 1) / nth;
|
|
720
|
-
int64_t start = duty * ith;
|
|
721
|
-
int64_t end = start + duty;
|
|
722
|
-
if (end > tiles)
|
|
723
|
-
end = tiles;
|
|
724
|
-
for (int64_t job = start; job < end; ++job) {
|
|
725
|
-
int64_t ii = m0 + job / xtiles * RM;
|
|
726
|
-
int64_t jj = n0 + job % xtiles * RN;
|
|
727
|
-
__m256 Cv[RN][RM] = {};
|
|
728
|
-
for (int64_t l = 0; l < k; ++l)
|
|
729
|
-
for (int64_t j = 0; j < RN; ++j)
|
|
730
|
-
for (int64_t i = 0; i < RM; ++i) {
|
|
731
|
-
#if defined(__AVX2__)
|
|
732
|
-
__m256 udTmp = updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
|
|
733
|
-
load(A + lda * (ii + i) + l)),
|
|
734
|
-
_mm256_sign_epi8(load(B + ldb * (jj + j) + l),
|
|
735
|
-
load(A + lda * (ii + i) + l)));
|
|
736
|
-
#else
|
|
737
|
-
__m128i ali0 = load0(A + lda * (ii + i) + l);
|
|
738
|
-
__m128i ali1 = load1(A + lda * (ii + i) + l);
|
|
739
|
-
__m128i blj0 = load0(B + ldb * (jj + j) + l);
|
|
740
|
-
__m128i blj1 = load1(B + ldb * (jj + j) + l);
|
|
741
|
-
|
|
742
|
-
__m128i sepAA0 = _mm_sign_epi8(ali0, ali0);
|
|
743
|
-
__m128i sepAA1 = _mm_sign_epi8(ali1, ali1);
|
|
744
|
-
__m128i sepBA0 = _mm_sign_epi8(blj0, ali0);
|
|
745
|
-
__m128i sepBA1 = _mm_sign_epi8(blj1, ali1);
|
|
746
|
-
|
|
747
|
-
// updot
|
|
748
|
-
const __m128i oneFill = _mm_set1_epi16(1);
|
|
749
|
-
__m128i mad0 = _mm_maddubs_epi16(sepAA0, sepBA0);
|
|
750
|
-
__m128i mad1 = _mm_maddubs_epi16(sepAA1, sepBA1);
|
|
751
|
-
__m256 udTmp = _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_madd_epi16(oneFill, mad1), _mm_madd_epi16(oneFill, mad0)));
|
|
752
|
-
#endif
|
|
753
|
-
Cv[j][i] = madd(_mm256_set1_ps(unhalf(A[lda * (ii + i) + l].d) *
|
|
754
|
-
unhalf(B[ldb * (jj + j) + l].d)),
|
|
755
|
-
udTmp,
|
|
756
|
-
Cv[j][i]);
|
|
757
|
-
}
|
|
758
|
-
for (int64_t j = 0; j < RN; ++j)
|
|
759
|
-
for (int64_t i = 0; i < RM; ++i)
|
|
760
|
-
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
|
|
761
|
-
}
|
|
762
|
-
}
|
|
763
|
-
|
|
764
|
-
inline __m256i load(const block_q8_0 *b) {
|
|
765
|
-
return _mm256_loadu_si256((const __m256i *)b->qs);
|
|
766
|
-
}
|
|
767
|
-
|
|
768
|
-
inline __m128i load0(const block_q8_0 *b) {
|
|
769
|
-
return _mm_loadu_si128((const __m128i *)b->qs);
|
|
770
|
-
}
|
|
771
|
-
|
|
772
|
-
inline __m128i load1(const block_q8_0 *b) {
|
|
773
|
-
return _mm_loadu_si128(((const __m128i *)b->qs) + 1);
|
|
774
|
-
}
|
|
775
|
-
|
|
776
|
-
inline __m256i load(const block_q4_0 *b) {
|
|
777
|
-
return _mm256_sub_epi8(denibble(b->qs), _mm256_set1_epi8(8));
|
|
778
|
-
}
|
|
779
|
-
|
|
780
|
-
inline __m128i load0(const block_q4_0 *b) {
|
|
781
|
-
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
|
|
782
|
-
return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), x), _mm_set1_epi8(8));
|
|
783
|
-
}
|
|
784
|
-
|
|
785
|
-
inline __m128i load1(const block_q4_0 *b) {
|
|
786
|
-
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
|
|
787
|
-
return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
|
|
788
|
-
}
|
|
789
|
-
|
|
790
|
-
inline __m256 updot(__m256i u, __m256i s) {
|
|
791
|
-
__m256i res;
|
|
792
|
-
#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
|
|
793
|
-
res = _mm256_dpbusd_epi32(_mm256_setzero_si256(), u, s);
|
|
794
|
-
#else
|
|
795
|
-
res = _mm256_madd_epi16(_mm256_set1_epi16(1), _mm256_maddubs_epi16(u, s));
|
|
796
|
-
#endif
|
|
797
|
-
return _mm256_cvtepi32_ps(res);
|
|
798
|
-
}
|
|
799
|
-
|
|
800
|
-
static inline __m256i denibble(const uint8_t *p) {
|
|
801
|
-
__m128i x = _mm_loadu_si128((const __m128i *)p);
|
|
802
|
-
return _mm256_and_si256(_mm256_set1_epi8(15),
|
|
803
|
-
_mm256_insertf128_si256(_mm256_castsi128_si256(x),
|
|
804
|
-
_mm_srli_epi16(x, 4), 1));
|
|
805
|
-
}
|
|
806
|
-
|
|
807
|
-
const TA *const A;
|
|
808
|
-
const TB *const B;
|
|
809
|
-
TC *const C;
|
|
810
|
-
const int64_t k;
|
|
811
|
-
const int64_t lda;
|
|
812
|
-
const int64_t ldb;
|
|
813
|
-
const int64_t ldc;
|
|
814
|
-
const int ith;
|
|
815
|
-
const int nth;
|
|
816
|
-
};
|
|
817
|
-
#endif // __AVX__
|
|
818
|
-
|
|
819
|
-
} // namespace
|
|
820
|
-
|
|
821
|
-
/**
|
|
822
|
-
* Performs optimized matrix multiplication on CPU.
|
|
823
|
-
*
|
|
824
|
-
* This subroutine may compute C = Aᵀ * B with column major ordering.
|
|
825
|
-
* Despite its name, this isn't a generalized implementation. Work is
|
|
826
|
-
* only performed when a handwritten kernel is written and available.
|
|
827
|
-
* Otherwise the caller should fall back to a general matmul routine.
|
|
828
|
-
*
|
|
829
|
-
* For example, for single-threaded single-precision GEMM you can say
|
|
830
|
-
*
|
|
831
|
-
* llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc,
|
|
832
|
-
* 0, 1, GGML_TASK_TYPE_COMPUTE,
|
|
833
|
-
* GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32);
|
|
834
|
-
*
|
|
835
|
-
* @param m is rows in `A` and `C`
|
|
836
|
-
* @param n is cols in `B` and `C`
|
|
837
|
-
* @param k is cols in `A` and rows in `B`
|
|
838
|
-
* @param A is first input matrix (always transposed)
|
|
839
|
-
* @param lda is row stride of `A`
|
|
840
|
-
* @param B is second input matrix (never transposed)
|
|
841
|
-
* @param ldb is row stride of `B`
|
|
842
|
-
* @param C is input/output array of output matrices
|
|
843
|
-
* @param ldc is row stride of `C`
|
|
844
|
-
* @param ith is thread id (must be less than `nth`)
|
|
845
|
-
* @param nth is number of threads (must be greater than zero)
|
|
846
|
-
* @param task is GGML task type
|
|
847
|
-
* @param Atype is GGML data type of `A`
|
|
848
|
-
* @param Btype is GGML data type of `B`
|
|
849
|
-
* @param Ctype is GGML data type of `C`
|
|
850
|
-
* @return true if this function was able to service the matmul request
|
|
851
|
-
*/
|
|
852
|
-
bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
|
|
853
|
-
int64_t ldc, int ith, int nth, int task, int Atype, int Btype, int Ctype) {
|
|
854
|
-
|
|
855
|
-
assert(m >= 0);
|
|
856
|
-
assert(n >= 0);
|
|
857
|
-
assert(k >= 0);
|
|
858
|
-
assert(lda >= k);
|
|
859
|
-
assert(ldb >= k);
|
|
860
|
-
assert(ldc >= m);
|
|
861
|
-
assert(nth > 0);
|
|
862
|
-
assert(ith < nth);
|
|
863
|
-
|
|
864
|
-
if (Ctype != GGML_TYPE_F32)
|
|
865
|
-
return false;
|
|
866
|
-
|
|
867
|
-
switch (Atype) {
|
|
868
|
-
|
|
869
|
-
case GGML_TYPE_F32: {
|
|
870
|
-
if (Btype != GGML_TYPE_F32)
|
|
871
|
-
return false;
|
|
872
|
-
#if defined(__AVX512F__)
|
|
873
|
-
if (k % 16)
|
|
874
|
-
return false;
|
|
875
|
-
tinyBLAS<16, __m512, __m512, float, float, float> tb{
|
|
876
|
-
k, (const float *)A, lda,
|
|
877
|
-
(const float *)B, ldb,
|
|
878
|
-
(float *)C, ldc,
|
|
879
|
-
ith, nth};
|
|
880
|
-
tb.matmul(m, n, task);
|
|
881
|
-
return true;
|
|
882
|
-
#elif defined(__AVX__) || defined(__AVX2__)
|
|
883
|
-
if (k % 8)
|
|
884
|
-
return false;
|
|
885
|
-
tinyBLAS<8, __m256, __m256, float, float, float> tb{
|
|
886
|
-
k, (const float *)A, lda,
|
|
887
|
-
(const float *)B, ldb,
|
|
888
|
-
(float *)C, ldc,
|
|
889
|
-
ith, nth};
|
|
890
|
-
tb.matmul(m, n, task);
|
|
891
|
-
return true;
|
|
892
|
-
#elif defined(__ARM_NEON)
|
|
893
|
-
if (n < 4)
|
|
894
|
-
return false;
|
|
895
|
-
if (k % 4)
|
|
896
|
-
return false;
|
|
897
|
-
tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
|
|
898
|
-
k, (const float *)A, lda,
|
|
899
|
-
(const float *)B, ldb,
|
|
900
|
-
(float *)C, ldc,
|
|
901
|
-
ith, nth};
|
|
902
|
-
tb.matmul(m, n, task);
|
|
903
|
-
return true;
|
|
904
|
-
#else
|
|
905
|
-
return false;
|
|
906
|
-
#endif
|
|
907
|
-
}
|
|
908
|
-
|
|
909
|
-
case GGML_TYPE_F16: {
|
|
910
|
-
#if defined(__AVX512F__)
|
|
911
|
-
if (k % 16)
|
|
912
|
-
return false;
|
|
913
|
-
if (Btype != GGML_TYPE_F32)
|
|
914
|
-
return false;
|
|
915
|
-
tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
|
|
916
|
-
k, (const ggml_fp16_t *)A, lda,
|
|
917
|
-
(const float *)B, ldb,
|
|
918
|
-
(float *)C, ldc,
|
|
919
|
-
ith, nth};
|
|
920
|
-
tb.matmul(m, n, task);
|
|
921
|
-
return true;
|
|
922
|
-
#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
|
|
923
|
-
if (k % 8)
|
|
924
|
-
return false;
|
|
925
|
-
if (Btype != GGML_TYPE_F32)
|
|
926
|
-
return false;
|
|
927
|
-
tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
|
|
928
|
-
k, (const ggml_fp16_t *)A, lda,
|
|
929
|
-
(const float *)B, ldb,
|
|
930
|
-
(float *)C, ldc,
|
|
931
|
-
ith, nth};
|
|
932
|
-
tb.matmul(m, n, task);
|
|
933
|
-
return true;
|
|
934
|
-
#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
|
|
935
|
-
if (n < 8)
|
|
936
|
-
return false;
|
|
937
|
-
if (k % 8)
|
|
938
|
-
return false;
|
|
939
|
-
if (Btype != GGML_TYPE_F16)
|
|
940
|
-
return false;
|
|
941
|
-
tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
|
|
942
|
-
k, (const ggml_fp16_t *)A, lda,
|
|
943
|
-
(const ggml_fp16_t *)B, ldb,
|
|
944
|
-
(float *)C, ldc,
|
|
945
|
-
ith, nth};
|
|
946
|
-
tb.matmul(m, n, task);
|
|
947
|
-
return true;
|
|
948
|
-
#elif defined(__ARM_NEON) && !defined(_MSC_VER)
|
|
949
|
-
if (k % 4)
|
|
950
|
-
return false;
|
|
951
|
-
if (Btype != GGML_TYPE_F32)
|
|
952
|
-
return false;
|
|
953
|
-
tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
|
|
954
|
-
k, (const ggml_fp16_t *)A, lda,
|
|
955
|
-
(const float *)B, ldb,
|
|
956
|
-
(float *)C, ldc,
|
|
957
|
-
ith, nth};
|
|
958
|
-
tb.matmul(m, n, task);
|
|
959
|
-
return true;
|
|
960
|
-
#else
|
|
961
|
-
return false;
|
|
962
|
-
#endif
|
|
963
|
-
}
|
|
964
|
-
|
|
965
|
-
case GGML_TYPE_Q8_0: {
|
|
966
|
-
if (Btype != GGML_TYPE_Q8_0)
|
|
967
|
-
return false;
|
|
968
|
-
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
|
|
969
|
-
tinyBLAS_Q0_AVX<block_q8_0, block_q8_0, float> tb{
|
|
970
|
-
k, (const block_q8_0 *)A, lda,
|
|
971
|
-
(const block_q8_0 *)B, ldb,
|
|
972
|
-
(float *)C, ldc,
|
|
973
|
-
ith, nth};
|
|
974
|
-
tb.matmul(m, n, task);
|
|
975
|
-
return true;
|
|
976
|
-
#elif defined(__ARM_FEATURE_DOTPROD)
|
|
977
|
-
tinyBLAS_Q0_ARM<block_q8_0> tb{
|
|
978
|
-
k, (const block_q8_0 *)A, lda,
|
|
979
|
-
(const block_q8_0 *)B, ldb,
|
|
980
|
-
(float *)C, ldc,
|
|
981
|
-
ith, nth};
|
|
982
|
-
tb.matmul(m, n, task);
|
|
983
|
-
return true;
|
|
984
|
-
#else
|
|
985
|
-
return false;
|
|
986
|
-
#endif
|
|
987
|
-
}
|
|
988
|
-
|
|
989
|
-
case GGML_TYPE_Q4_0: {
|
|
990
|
-
if (Btype != GGML_TYPE_Q8_0)
|
|
991
|
-
return false;
|
|
992
|
-
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
|
|
993
|
-
tinyBLAS_Q0_AVX<block_q4_0, block_q8_0, float> tb{
|
|
994
|
-
k, (const block_q4_0 *)A, lda,
|
|
995
|
-
(const block_q8_0 *)B, ldb,
|
|
996
|
-
(float *)C, ldc,
|
|
997
|
-
ith, nth};
|
|
998
|
-
tb.matmul(m, n, task);
|
|
999
|
-
return true;
|
|
1000
|
-
#elif defined(__ARM_FEATURE_DOTPROD)
|
|
1001
|
-
tinyBLAS_Q0_ARM<block_q4_0> tb{
|
|
1002
|
-
k, (const block_q4_0 *)A, lda,
|
|
1003
|
-
(const block_q8_0 *)B, ldb,
|
|
1004
|
-
(float *)C, ldc,
|
|
1005
|
-
ith, nth};
|
|
1006
|
-
tb.matmul(m, n, task);
|
|
1007
|
-
return true;
|
|
1008
|
-
#else
|
|
1009
|
-
return false;
|
|
1010
|
-
#endif
|
|
1011
|
-
}
|
|
1012
|
-
|
|
1013
|
-
default:
|
|
1014
|
-
return false;
|
|
1015
|
-
}
|
|
1016
|
-
|
|
1017
|
-
(void)m;
|
|
1018
|
-
(void)n;
|
|
1019
|
-
(void)k;
|
|
1020
|
-
(void)A;
|
|
1021
|
-
(void)lda;
|
|
1022
|
-
(void)B;
|
|
1023
|
-
(void)ldb;
|
|
1024
|
-
(void)C;
|
|
1025
|
-
(void)ldc;
|
|
1026
|
-
(void)ith;
|
|
1027
|
-
(void)nth;
|
|
1028
|
-
(void)task;
|
|
1029
|
-
(void)Atype;
|
|
1030
|
-
(void)Btype;
|
|
1031
|
-
(void)Ctype;
|
|
1032
|
-
}
|