From 1115a966a4921a1e4850a2267554074084b66f7d Mon Sep 17 00:00:00 2001
From: Viet-Hoa Do <viet-hoa.do@arm.com>
Date: Thu, 30 May 2024 15:26:04 +0100
Subject: [PATCH] Restructure the project

* The library is move into kai directory (previously src).
* All microkernels are in kai/ukernels directory.
* Packing kernels for matmul will be in kai/ukernels/matmul/pack
  directory.

Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
---
 CMakeLists.txt                                | 27 ++++++++-----------
 {src => kai}/kai_common.h                     |  0
 ..._f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c |  2 +-
 ..._f16_f16_f16p16x1biasf16_6x16x8_neon_mla.h |  0
 ...ai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c |  2 +-
 ...ai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.h |  0
 ...ai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c |  2 +-
 ...ai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.h |  0
 ...2_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c |  2 +-
 ...2_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.h |  0
 ...2_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c |  2 +-
 ...2_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.h |  0
 ...2_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c |  2 +-
 ...2_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.h |  0
 ...2_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c |  2 +-
 ...2_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.h |  0
 ...tmul_clamp_f32_qai8dxp_qsi4cxp_interface.h |  0
 .../pack}/kai_lhs_quant_pack_qai8dxp_f32.c    |  2 +-
 .../pack}/kai_lhs_quant_pack_qai8dxp_f32.h    |  0
 ...hs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c |  2 +-
 ...hs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h |  0
 .../kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c     |  2 +-
 .../kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.h     |  0
 test/common/compare.cpp                       |  2 +-
 test/common/data_format.cpp                   |  2 +-
 test/common/data_type.cpp                     |  2 +-
 test/common/int4.cpp                          |  2 +-
 test/common/matrix_portion.cpp                |  2 +-
 test/common/printer.cpp                       |  2 +-
 test/reference/binary_elementwise.cpp         |  2 +-
 test/reference/cast.cpp                       |  2 +-
 test/reference/fill.cpp                       |  2 +-
 test/reference/matmul.cpp                     |  2 +-
 test/reference/pack.cpp                       |  2 +-
 test/reference/quantize.cpp                   |  2 +-
 test/reference/reduce.cpp                     |  2 +-
 test/reference/transpose.cpp                  |  2 +-
 test/tests/matmul_test.cpp                    |  6 ++---
 38 files changed, 38 insertions(+), 43 deletions(-)
 rename {src => kai}/kai_common.h (100%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c (99%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.h (100%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c (99%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.h (100%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c (99%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.h (100%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c (99%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.h (100%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c (99%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.h (100%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c (99%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.h (100%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c (99%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.h (100%)
 rename {src => kai/ukernels}/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp_qsi4cxp_interface.h (100%)
 rename {src/matmul => kai/ukernels/matmul/pack}/kai_lhs_quant_pack_qai8dxp_f32.c (99%)
 rename {src/matmul => kai/ukernels/matmul/pack}/kai_lhs_quant_pack_qai8dxp_f32.h (100%)
 rename {src/matmul => kai/ukernels/matmul/pack}/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c (99%)
 rename {src/matmul => kai/ukernels/matmul/pack}/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h (100%)
 rename {src/matmul => kai/ukernels/matmul/pack}/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c (99%)
 rename {src/matmul => kai/ukernels/matmul/pack}/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.h (100%)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0ff56f89..822c1137 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -57,25 +57,25 @@ set(KLEIDIAI_WARNING_FLAGS
 )
 
 set(KLEIDIAI_FILES_NEON
-    src/matmul/kai_lhs_quant_pack_qai8dxp_f32.c
-    src/matmul/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c
+    kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
+    kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c
 )
 
 set(KLEIDIAI_FILES_NEON_FP16
-    src/matmul/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
-    src/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c
+    kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
+    kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c
 )
 
 set(KLEIDIAI_FILES_NEON_DOTPROD
-    src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c
-    src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c
+    kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c
+    kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c
 )
 
 set(KLEIDIAI_FILES_NEON_I8MM
-    src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c
-    src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c
-    src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c
-    src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c
+    kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c
+    kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c
+    kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c
+    kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c
 )
 
 add_library(kleidiai
@@ -86,8 +86,7 @@ add_library(kleidiai
 )
 
 target_include_directories(kleidiai
-    PUBLIC  src
-    PRIVATE .
+    PUBLIC  .
 )
 
 target_compile_options(kleidiai
@@ -142,10 +141,6 @@ if(KLEIDIAI_BUILD_TESTS)
         test/tests/matmul_test.cpp
     )
 
-    target_include_directories(kleidiai_test
-        PRIVATE .
-    )
-
     target_compile_options(kleidiai_test
         PRIVATE ${KLEIDIAI_WARNING_FLAGS}
         PRIVATE -march=armv8.2-a+fp16+bf16
diff --git a/src/kai_common.h b/kai/kai_common.h
similarity index 100%
rename from src/kai_common.h
rename to kai/kai_common.h
diff --git a/src/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c b/kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c
similarity index 99%
rename from src/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c
rename to kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c
index c4996ebd..8c848afd 100644
--- a/src/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c
+++ b/kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.c
@@ -13,7 +13,7 @@
 #include <stddef.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_mr = 6;
 static const size_t kai_nr = 16;
diff --git a/src/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.h b/kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.h
similarity index 100%
rename from src/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.h
rename to kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.h
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c
similarity index 99%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c
index 4d47ef1e..fa9a3f5c 100644
--- a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c
+++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.c
@@ -11,7 +11,7 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_m_step = 1;
 static const size_t kai_n_step = 4;
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.h b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.h
similarity index 100%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.h
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp4x8_1x4x32_neon_dotprod.h
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c
similarity index 99%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c
index 6bff095a..17707488 100644
--- a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c
+++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.c
@@ -11,7 +11,7 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_m_step = 1;
 static const size_t kai_n_step = 8;
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.h b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.h
similarity index 100%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.h
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp1x8_qsi4cxp8x8_1x8x32_neon_dotprod.h
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c
similarity index 99%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c
index 45f43170..ebcaaf54 100644
--- a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c
+++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.c
@@ -11,7 +11,7 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_m_step = 4;
 static const size_t kai_n_step = 4;
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.h b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.h
similarity index 100%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.h
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_4x4x32_neon_i8mm.h
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c
similarity index 99%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c
index 293fad29..ac9ec40e 100644
--- a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c
+++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.c
@@ -11,7 +11,7 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_m_step = 8;
 static const size_t kai_n_step = 4;
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.h b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.h
similarity index 100%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.h
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x8_8x4x32_neon_i8mm.h
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c
similarity index 99%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c
index 06adc07a..93c1a350 100644
--- a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c
+++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.c
@@ -11,7 +11,7 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_m_step = 4;
 static const size_t kai_n_step = 8;
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.h b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.h
similarity index 100%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.h
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.h
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c
similarity index 99%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c
index 9d8153a9..2b2b1e09 100644
--- a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c
+++ b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.c
@@ -11,7 +11,7 @@
 #include <arm_neon.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_m_step = 8;
 static const size_t kai_n_step = 8;
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.h b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.h
similarity index 100%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.h
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_8x8x32_neon_i8mm.h
diff --git a/src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp_qsi4cxp_interface.h b/kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp_qsi4cxp_interface.h
similarity index 100%
rename from src/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp_qsi4cxp_interface.h
rename to kai/ukernels/matmul/matmul_clamp_f32_qai8dxp_qsi4cxp/kai_matmul_clamp_f32_qai8dxp_qsi4cxp_interface.h
diff --git a/src/matmul/kai_lhs_quant_pack_qai8dxp_f32.c b/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
similarity index 99%
rename from src/matmul/kai_lhs_quant_pack_qai8dxp_f32.c
rename to kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
index 6e618ccc..ffb09949 100644
--- a/src/matmul/kai_lhs_quant_pack_qai8dxp_f32.c
+++ b/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.c
@@ -10,7 +10,7 @@
 #include <math.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_num_bytes_per_multiplier = sizeof(float);
 static const size_t kai_num_bytes_per_offset = sizeof(int32_t);
diff --git a/src/matmul/kai_lhs_quant_pack_qai8dxp_f32.h b/kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.h
similarity index 100%
rename from src/matmul/kai_lhs_quant_pack_qai8dxp_f32.h
rename to kai/ukernels/matmul/pack/kai_lhs_quant_pack_qai8dxp_f32.h
diff --git a/src/matmul/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
similarity index 99%
rename from src/matmul/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
rename to kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
index 4bf04303..50c04ad3 100644
--- a/src/matmul/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.c
@@ -11,7 +11,7 @@
 #include <stddef.h>
 #include <stdint.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_nr = 16;
 static const size_t kai_kr = 1;
diff --git a/src/matmul/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h b/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h
similarity index 100%
rename from src/matmul/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h
rename to kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h
diff --git a/src/matmul/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c
similarity index 99%
rename from src/matmul/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c
rename to kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c
index aed8a5f4..41d3fcdb 100644
--- a/src/matmul/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c
+++ b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.c
@@ -10,7 +10,7 @@
 #include <stdint.h>
 #include <string.h>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 static const size_t kai_num_bytes_sum_rhs = sizeof(int32_t);
 static const size_t kai_num_bytes_multiplier_rhs = sizeof(float);
diff --git a/src/matmul/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.h b/kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.h
similarity index 100%
rename from src/matmul/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.h
rename to kai/ukernels/matmul/pack/kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0.h
diff --git a/test/common/compare.cpp b/test/common/compare.cpp
index 9dcfb104..91b6bc66 100644
--- a/test/common/compare.cpp
+++ b/test/common/compare.cpp
@@ -12,7 +12,7 @@
 #include <tuple>
 #include <type_traits>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/bfloat16.hpp"
 #include "test/common/data_format.hpp"
 #include "test/common/data_type.hpp"
diff --git a/test/common/data_format.cpp b/test/common/data_format.cpp
index a69a085d..a16269e9 100644
--- a/test/common/data_format.cpp
+++ b/test/common/data_format.cpp
@@ -9,7 +9,7 @@
 #include <cstddef>
 #include <cstdint>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/data_type.hpp"
 #include "test/reference/round.hpp"
 
diff --git a/test/common/data_type.cpp b/test/common/data_type.cpp
index 95209750..44f594af 100644
--- a/test/common/data_type.cpp
+++ b/test/common/data_type.cpp
@@ -9,7 +9,7 @@
 #include <cstddef>
 #include <cstdint>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 
 namespace kai::test {
 
diff --git a/test/common/int4.cpp b/test/common/int4.cpp
index ae0a5998..5f8afef5 100644
--- a/test/common/int4.cpp
+++ b/test/common/int4.cpp
@@ -9,7 +9,7 @@
 #include <cstdint>
 #include <tuple>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 
 namespace kai::test {
 
diff --git a/test/common/matrix_portion.cpp b/test/common/matrix_portion.cpp
index e7c42bf1..6b975ec7 100644
--- a/test/common/matrix_portion.cpp
+++ b/test/common/matrix_portion.cpp
@@ -9,7 +9,7 @@
 #include <algorithm>
 #include <cstddef>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/rect.hpp"
 #include "test/reference/round.hpp"
 
diff --git a/test/common/printer.cpp b/test/common/printer.cpp
index 320f7244..7d7144d8 100644
--- a/test/common/printer.cpp
+++ b/test/common/printer.cpp
@@ -9,7 +9,7 @@
 #include <ostream>
 #include <string_view>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/bfloat16.hpp"
 #include "test/common/data_format.hpp"
 #include "test/common/data_type.hpp"
diff --git a/test/reference/binary_elementwise.cpp b/test/reference/binary_elementwise.cpp
index c5c61a66..d5abd26c 100644
--- a/test/reference/binary_elementwise.cpp
+++ b/test/reference/binary_elementwise.cpp
@@ -11,7 +11,7 @@
 #include <cstdint>
 #include <vector>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/data_type.hpp"
 #include "test/common/float16.hpp"
 #include "test/common/int4.hpp"
diff --git a/test/reference/cast.cpp b/test/reference/cast.cpp
index 1eae56df..21486c05 100644
--- a/test/reference/cast.cpp
+++ b/test/reference/cast.cpp
@@ -8,7 +8,7 @@
 #include <cstdint>
 #include <vector>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/bfloat16.hpp"
 #include "test/common/data_type.hpp"
 #include "test/common/memory.hpp"
diff --git a/test/reference/fill.cpp b/test/reference/fill.cpp
index 238397c0..d58ab6ce 100644
--- a/test/reference/fill.cpp
+++ b/test/reference/fill.cpp
@@ -13,7 +13,7 @@
 #include <type_traits>
 #include <vector>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/bfloat16.hpp"
 #include "test/common/data_format.hpp"
 #include "test/common/data_type.hpp"
diff --git a/test/reference/matmul.cpp b/test/reference/matmul.cpp
index 803fa23b..984982a2 100644
--- a/test/reference/matmul.cpp
+++ b/test/reference/matmul.cpp
@@ -10,7 +10,7 @@
 #include <cstdint>
 #include <vector>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/data_format.hpp"
 #include "test/common/data_type.hpp"
 #include "test/common/float16.hpp"
diff --git a/test/reference/pack.cpp b/test/reference/pack.cpp
index a61efff8..261d7a53 100644
--- a/test/reference/pack.cpp
+++ b/test/reference/pack.cpp
@@ -14,7 +14,7 @@
 #include <limits>
 #include <vector>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/data_format.hpp"
 #include "test/common/data_type.hpp"
 #include "test/reference/quantize.hpp"
diff --git a/test/reference/quantize.cpp b/test/reference/quantize.cpp
index 141892b2..213328d5 100644
--- a/test/reference/quantize.cpp
+++ b/test/reference/quantize.cpp
@@ -13,7 +13,7 @@
 #include <tuple>
 #include <vector>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/data_type.hpp"
 #include "test/common/int4.hpp"
 #include "test/common/memory.hpp"
diff --git a/test/reference/reduce.cpp b/test/reference/reduce.cpp
index dff4cf69..0e83b9bf 100644
--- a/test/reference/reduce.cpp
+++ b/test/reference/reduce.cpp
@@ -10,7 +10,7 @@
 #include <cstdint>
 #include <vector>
 
-#include "src/kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/data_format.hpp"
 #include "test/common/data_type.hpp"
 #include "test/common/int4.hpp"
diff --git a/test/reference/transpose.cpp b/test/reference/transpose.cpp
index bee2182a..95cbc8e2 100644
--- a/test/reference/transpose.cpp
+++ b/test/reference/transpose.cpp
@@ -11,7 +11,7 @@
 #include <cstring>
 #include <vector>
 
-#include "kai_common.h"
+#include "kai/kai_common.h"
 #include "test/common/data_type.hpp"
 
 namespace kai::test {
diff --git a/test/tests/matmul_test.cpp b/test/tests/matmul_test.cpp
index 544d17e9..65c5f603 100644
--- a/test/tests/matmul_test.cpp
+++ b/test/tests/matmul_test.cpp
@@ -20,9 +20,9 @@
 #include <utility>
 #include <vector>
 
-#include "src/kai_common.h"
-#include "src/matmul/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h"
-#include "src/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.h"
+#include "kai/kai_common.h"
+#include "kai/ukernels/matmul/matmul_clamp_f16_f16_f16p/kai_matmul_clamp_f16_f16_f16p16x1biasf16_6x16x8_neon_mla.h"
+#include "kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f16p16x1biasf16_f16_f16_neon.h"
 #include "test/common/compare.hpp"
 #include "test/common/data_format.hpp"
 #include "test/common/data_type.hpp"
-- 
GitLab