mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-06-27 12:05:03 +00:00
ggml-cpu : split arch-specific implementations (#13892)
* move ggml-cpu-aarch64 to repack * split quantize_row_q8_0/1 * split helper functions * split ggml_vec_dot_q4_0_q8_0 * split ggml_vec_dot_q4_1_q8_1 * split ggml_vec_dot_q5_0_q8_0 * split ggml_vec_dot_q5_1_q8_1 * split ggml_vec_dot_q8_0_q8_0 * split ggml_vec_dot_tq1_0_q8_K * split ggml_vec_dot_tq2_0_q8_K * split ggml_vec_dot_q2_K_q8_K * split ggml_vec_dot_q3_K_q8_K * split ggml_vec_dot_q4_K_q8_K * split ggml_vec_dot_q5_K_q8_K * split ggml_vec_dot_q6_K_q8_K * split ggml_vec_dot_iq2_xxs_q8_K * split ggml_vec_dot_iq2_xs_q8_K * split ggml_vec_dot_iq2_s_q8_K * split ggml_vec_dot_iq3_xxs_q8_K * split ggml_vec_dot_iq3_s_q8_K * split ggml_vec_dot_iq1_s_q8_K * split ggml_vec_dot_iq1_m_q8_K * split ggml_vec_dot_iq4_nl_q8_0 * split ggml_vec_dot_iq4_xs_q8_K * fix typos * fix missing prototypes * rename ggml-cpu-quants.c * rename ggml-cpu-traits * rename arm folder * move cpu-feats-x86.cpp * rename ggml-cpu-hbm * update arm detection macro in quants.c * move iq quant tables * split ggml_quantize_mat_q8_0/K * split ggml_gemv_* * split ggml_gemm_* * rename namespace aarch64 to repack * use weak aliases to replace test macros * rename GGML_CPU_AARCH64 to GGML_CPU_REPACK * rename more aarch64 to repack * clean up rebase leftover * fix compilation errors * remove trailing spaces * try to fix clang compilation errors * try to fix clang compilation errors again * try to fix clang compilation errors, 3rd attempt * try to fix clang compilation errors, 4th attempt * try to fix clang compilation errors, 5th attempt * try to fix clang compilation errors, 6th attempt * try to fix clang compilation errors, 7th attempt * try to fix clang compilation errors, 8th attempt * try to fix clang compilation errors, 9th attempt * more cleanup * fix compilation errors * fix apple targets * fix a typo in arm version of ggml_vec_dot_q4_K_q8_K Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
xctan
GitHub
4
Makefile
4
Makefile
@ -367,7 +367,7 @@ ifdef LLAMA_SERVER_SSL
|
||||
endif
|
||||
|
||||
ifndef GGML_NO_CPU_AARCH64
|
||||
MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
|
||||
MK_CPPFLAGS += -DGGML_USE_CPU_REPACK
|
||||
endif
|
||||
|
||||
# warnings
|
||||
@ -970,7 +970,7 @@ OBJ_GGML = \
|
||||
$(DIR_GGML)/src/ggml-threading.o \
|
||||
$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
|
||||
$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
|
||||
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
|
||||
$(DIR_GGML)/src/ggml-cpu/repack.o \
|
||||
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
|
||||
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
|
||||
$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \
|
||||
|
||||
@ -105,7 +105,7 @@ message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
|
||||
message(DEBUG "INS_ENB : ${INS_ENB}")
|
||||
|
||||
option(GGML_CPU_HBM "ggml: use memkind for CPU HBM" OFF)
|
||||
option(GGML_CPU_AARCH64 "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
|
||||
option(GGML_CPU_REPACK "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
|
||||
option(GGML_CPU_KLEIDIAI "ggml: use KleidiAI optimized kernels if applicable" OFF)
|
||||
option(GGML_SSE42 "ggml: enable SSE 4.2" ${INS_ENB})
|
||||
option(GGML_AVX "ggml: enable AVX" ${INS_ENB})
|
||||
|
||||
@ -1074,6 +1074,10 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
|
||||
0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
|
||||
GGML_TABLE_END()
|
||||
|
||||
GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
|
||||
-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
|
||||
GGML_TABLE_END()
|
||||
|
||||
#define NGRID_IQ1S 2048
|
||||
#define IQ1S_DELTA 0.125f
|
||||
#define IQ1M_DELTA 0.125f
|
||||
|
||||
@ -10,14 +10,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
list (APPEND GGML_CPU_SOURCES
|
||||
ggml-cpu/ggml-cpu.c
|
||||
ggml-cpu/ggml-cpu.cpp
|
||||
ggml-cpu/ggml-cpu-aarch64.cpp
|
||||
ggml-cpu/ggml-cpu-aarch64.h
|
||||
ggml-cpu/ggml-cpu-hbm.cpp
|
||||
ggml-cpu/ggml-cpu-hbm.h
|
||||
ggml-cpu/ggml-cpu-quants.c
|
||||
ggml-cpu/ggml-cpu-quants.h
|
||||
ggml-cpu/ggml-cpu-traits.cpp
|
||||
ggml-cpu/ggml-cpu-traits.h
|
||||
ggml-cpu/repack.cpp
|
||||
ggml-cpu/repack.h
|
||||
ggml-cpu/hbm.cpp
|
||||
ggml-cpu/hbm.h
|
||||
ggml-cpu/quants.c
|
||||
ggml-cpu/quants.h
|
||||
ggml-cpu/traits.cpp
|
||||
ggml-cpu/traits.h
|
||||
ggml-cpu/amx/amx.cpp
|
||||
ggml-cpu/amx/amx.h
|
||||
ggml-cpu/amx/mmq.cpp
|
||||
@ -84,6 +84,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
|
||||
if (GGML_SYSTEM_ARCH STREQUAL "ARM")
|
||||
message(STATUS "ARM detected")
|
||||
list(APPEND GGML_CPU_SOURCES
|
||||
ggml-cpu/arch/arm/quants.c
|
||||
ggml-cpu/arch/arm/repack.cpp
|
||||
)
|
||||
|
||||
if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
|
||||
message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
|
||||
else()
|
||||
@ -167,6 +172,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
endif()
|
||||
elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
|
||||
message(STATUS "x86 detected")
|
||||
list(APPEND GGML_CPU_SOURCES
|
||||
ggml-cpu/arch/x86/quants.c
|
||||
ggml-cpu/arch/x86/repack.cpp
|
||||
)
|
||||
|
||||
if (MSVC)
|
||||
# instruction set detection for MSVC only
|
||||
if (GGML_NATIVE)
|
||||
@ -302,7 +312,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
# Since multiple variants of the CPU backend may be included in the same
|
||||
# build, using set_source_files_properties() to set the arch flags is not possible
|
||||
set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
|
||||
add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
|
||||
add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/x86/cpu-feats.cpp)
|
||||
target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
|
||||
target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
|
||||
target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
|
||||
@ -311,6 +321,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
endif()
|
||||
elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
|
||||
message(STATUS "PowerPC detected")
|
||||
list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c)
|
||||
if (GGML_NATIVE)
|
||||
if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
|
||||
file(READ "/proc/cpuinfo" POWER10_M)
|
||||
@ -338,6 +349,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
endif()
|
||||
elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64")
|
||||
message(STATUS "loongarch64 detected")
|
||||
list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c)
|
||||
|
||||
list(APPEND ARCH_FLAGS -march=loongarch64)
|
||||
if (GGML_LASX)
|
||||
list(APPEND ARCH_FLAGS -mlasx)
|
||||
@ -347,6 +360,10 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
endif()
|
||||
elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
|
||||
message(STATUS "riscv64 detected")
|
||||
list(APPEND GGML_CPU_SOURCES
|
||||
ggml-cpu/arch/riscv/quants.c
|
||||
ggml-cpu/arch/riscv/repack.cpp
|
||||
)
|
||||
if (GGML_RVV)
|
||||
if (GGML_XTHEADVECTOR)
|
||||
list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
|
||||
@ -358,6 +375,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
endif()
|
||||
elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
|
||||
message(STATUS "s390x detected")
|
||||
list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
|
||||
file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
|
||||
string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
|
||||
|
||||
@ -381,12 +399,16 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
|
||||
if (GGML_VXE)
|
||||
list(APPEND ARCH_FLAGS -mvx -mzvector)
|
||||
endif()
|
||||
elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
|
||||
message(STATUS "Wasm detected")
|
||||
list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
|
||||
else()
|
||||
message(STATUS "Unknown architecture")
|
||||
message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
|
||||
list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
|
||||
endif()
|
||||
|
||||
if (GGML_CPU_AARCH64)
|
||||
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
|
||||
if (GGML_CPU_REPACK)
|
||||
target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
|
||||
endif()
|
||||
|
||||
if (GGML_CPU_KLEIDIAI)
|
||||
|
||||
@ -5,7 +5,7 @@
|
||||
#include "ggml-backend.h"
|
||||
#include "ggml-impl.h"
|
||||
#include "ggml-cpu.h"
|
||||
#include "ggml-cpu-traits.h"
|
||||
#include "traits.h"
|
||||
|
||||
#if defined(__gnu_linux__)
|
||||
#include <sys/syscall.h>
|
||||
|
||||
@ -8,7 +8,7 @@
|
||||
#include "mmq.h"
|
||||
#include "ggml-impl.h"
|
||||
#include "ggml-cpu-impl.h"
|
||||
#include "ggml-cpu-quants.h"
|
||||
#include "quants.h"
|
||||
#include "ggml-quants.h"
|
||||
#include <algorithm>
|
||||
#include <type_traits>
|
||||
|
||||
4113
ggml/src/ggml-cpu/arch/arm/quants.c
Normal file
4113
ggml/src/ggml-cpu/arch/arm/quants.c
Normal file
File diff suppressed because it is too large
Load Diff
2174
ggml/src/ggml-cpu/arch/arm/repack.cpp
Normal file
2174
ggml/src/ggml-cpu/arch/arm/repack.cpp
Normal file
File diff suppressed because it is too large
Load Diff
2638
ggml/src/ggml-cpu/arch/loongarch/quants.c
Normal file
2638
ggml/src/ggml-cpu/arch/loongarch/quants.c
Normal file
File diff suppressed because it is too large
Load Diff
2731
ggml/src/ggml-cpu/arch/powerpc/quants.c
Normal file
2731
ggml/src/ggml-cpu/arch/powerpc/quants.c
Normal file
File diff suppressed because it is too large
Load Diff
2068
ggml/src/ggml-cpu/arch/riscv/quants.c
Normal file
2068
ggml/src/ggml-cpu/arch/riscv/quants.c
Normal file
File diff suppressed because it is too large
Load Diff
396
ggml/src/ggml-cpu/arch/riscv/repack.cpp
Normal file
396
ggml/src/ggml-cpu/arch/riscv/repack.cpp
Normal file
@ -0,0 +1,396 @@
|
||||
#define GGML_COMMON_IMPL_CPP
|
||||
#define GGML_COMMON_DECL_CPP
|
||||
#include "ggml-common.h"
|
||||
#include "ggml-backend-impl.h"
|
||||
|
||||
#include "ggml-impl.h"
|
||||
#include "ggml-cpu.h"
|
||||
#include "ggml-cpu-impl.h"
|
||||
#include "traits.h"
|
||||
|
||||
#include <cmath>
|
||||
#include <cstring>
|
||||
#include <cassert>
|
||||
#include <cstdlib> // for qsort
|
||||
#include <cstdio> // for GGML_ASSERT
|
||||
|
||||
#define GGML_CPU_CLANG_WORKAROUND
|
||||
#include "../../repack.h"
|
||||
|
||||
#if defined(__GNUC__)
|
||||
#pragma GCC diagnostic ignored "-Woverlength-strings"
|
||||
#endif
|
||||
|
||||
#define UNUSED GGML_UNUSED
|
||||
|
||||
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
const int qk = QK8_0;
|
||||
const int nb = n / qk;
|
||||
const int ncols_interleaved = 8;
|
||||
const int blocklen = 8;
|
||||
|
||||
assert (n % qk == 0);
|
||||
assert (nc % ncols_interleaved == 0);
|
||||
|
||||
UNUSED(s);
|
||||
UNUSED(bs);
|
||||
UNUSED(vx);
|
||||
UNUSED(vy);
|
||||
UNUSED(nr);
|
||||
UNUSED(nc);
|
||||
UNUSED(nb);
|
||||
UNUSED(ncols_interleaved);
|
||||
UNUSED(blocklen);
|
||||
|
||||
#if defined __riscv_v
|
||||
if (__riscv_vlenb() >= QK4_0) {
|
||||
const size_t vl = QK4_0;
|
||||
|
||||
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
|
||||
|
||||
vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
for (int l = 0; l < nb; l++) {
|
||||
const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
|
||||
const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
|
||||
const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
|
||||
const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
|
||||
|
||||
const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
|
||||
const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
|
||||
const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
|
||||
const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
|
||||
const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
|
||||
const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
|
||||
const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
|
||||
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
// vector version needs Zvfhmin extension
|
||||
const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
|
||||
const float b_scales[8] = {
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[0]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[1]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[2]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[3]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[4]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[5]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[6]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[7])
|
||||
};
|
||||
const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
|
||||
sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
__riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
|
||||
}
|
||||
return;
|
||||
}
|
||||
|
||||
#endif
|
||||
{
|
||||
float sumf[8];
|
||||
int sumi;
|
||||
|
||||
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
|
||||
|
||||
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
|
||||
for (int l = 0; l < nb; l++) {
|
||||
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
sumi = 0;
|
||||
for (int i = 0; i < blocklen; ++i) {
|
||||
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
|
||||
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
|
||||
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
|
||||
}
|
||||
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
|
||||
}
|
||||
}
|
||||
}
|
||||
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
|
||||
const int qk = QK8_0;
|
||||
const int nb = n / qk;
|
||||
const int ncols_interleaved = 8;
|
||||
const int blocklen = 8;
|
||||
|
||||
assert (n % qk == 0);
|
||||
assert (nr % 4 == 0);
|
||||
assert (nc % ncols_interleaved == 0);
|
||||
|
||||
UNUSED(s);
|
||||
UNUSED(bs);
|
||||
UNUSED(vx);
|
||||
UNUSED(vy);
|
||||
UNUSED(nr);
|
||||
UNUSED(nc);
|
||||
UNUSED(nb);
|
||||
UNUSED(ncols_interleaved);
|
||||
UNUSED(blocklen);
|
||||
|
||||
#if defined __riscv_v
|
||||
if (__riscv_vlenb() >= QK4_0) {
|
||||
const size_t vl = QK4_0;
|
||||
|
||||
for (int y = 0; y < nr / 4; y++) {
|
||||
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
|
||||
vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
|
||||
for (int l = 0; l < nb; l++) {
|
||||
const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
|
||||
const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
|
||||
const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
|
||||
const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
|
||||
const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
|
||||
const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
|
||||
const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
|
||||
|
||||
// vector version needs Zvfhmin extension
|
||||
const float a_scales[4] = {
|
||||
GGML_FP16_TO_FP32(a_ptr[l].d[0]),
|
||||
GGML_FP16_TO_FP32(a_ptr[l].d[1]),
|
||||
GGML_FP16_TO_FP32(a_ptr[l].d[2]),
|
||||
GGML_FP16_TO_FP32(a_ptr[l].d[3])
|
||||
};
|
||||
const float b_scales[8] = {
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[0]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[1]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[2]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[3]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[4]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[5]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[6]),
|
||||
GGML_FP16_TO_FP32(b_ptr[l].d[7])
|
||||
};
|
||||
const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
|
||||
|
||||
const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
|
||||
const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
|
||||
const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
|
||||
const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
vint16m4_t sumi_l0;
|
||||
{
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
sumi_l0 = sumi_hi_m;
|
||||
}
|
||||
|
||||
{
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
|
||||
sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
|
||||
const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
|
||||
const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
|
||||
const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
|
||||
const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
vint16m4_t sumi_l1;
|
||||
{
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
sumi_l1 = sumi_hi_m;
|
||||
}
|
||||
|
||||
{
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
|
||||
sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
|
||||
const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
|
||||
const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
|
||||
const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
|
||||
const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
vint16m4_t sumi_l2;
|
||||
{
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
sumi_l2 = sumi_hi_m;
|
||||
}
|
||||
|
||||
{
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
|
||||
sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
|
||||
const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
|
||||
const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
|
||||
const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
|
||||
const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
|
||||
__asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
|
||||
vint16m4_t sumi_l3;
|
||||
{
|
||||
const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
|
||||
const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
|
||||
const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
|
||||
const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
|
||||
const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
|
||||
const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
|
||||
const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
|
||||
|
||||
sumi_l3 = sumi_hi_m;
|
||||
}
|
||||
|
||||
{
|
||||
const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
|
||||
const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
|
||||
const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
|
||||
const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
|
||||
const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
|
||||
const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
|
||||
const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
|
||||
const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
|
||||
const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
|
||||
const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
|
||||
const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
|
||||
const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
|
||||
const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
|
||||
|
||||
const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
|
||||
sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
|
||||
}
|
||||
}
|
||||
__riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
|
||||
__riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
|
||||
__riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
|
||||
__riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
|
||||
}
|
||||
}
|
||||
|
||||
return;
|
||||
}
|
||||
|
||||
#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
|
||||
float sumf[4][8];
|
||||
int sumi;
|
||||
|
||||
for (int y = 0; y < nr / 4; y++) {
|
||||
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
|
||||
for (int x = 0; x < nc / ncols_interleaved; x++) {
|
||||
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
|
||||
}
|
||||
for (int l = 0; l < nb; l++) {
|
||||
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++) {
|
||||
sumi = 0;
|
||||
for (int i = 0; i < blocklen; ++i) {
|
||||
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
|
||||
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
|
||||
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
|
||||
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
|
||||
}
|
||||
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
for (int m = 0; m < 4; m++) {
|
||||
for (int j = 0; j < ncols_interleaved; j++)
|
||||
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
1299
ggml/src/ggml-cpu/arch/s390/quants.c
Normal file
1299
ggml/src/ggml-cpu/arch/s390/quants.c
Normal file
File diff suppressed because it is too large
Load Diff
1480
ggml/src/ggml-cpu/arch/wasm/quants.c
Normal file
1480
ggml/src/ggml-cpu/arch/wasm/quants.c
Normal file
File diff suppressed because it is too large
Load Diff
4310
ggml/src/ggml-cpu/arch/x86/quants.c
Normal file
4310
ggml/src/ggml-cpu/arch/x86/quants.c
Normal file
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@ -1,7 +1,7 @@
|
||||
#pragma once
|
||||
|
||||
#include "ggml.h"
|
||||
#include "ggml-cpu-traits.h"
|
||||
#include "traits.h"
|
||||
#include "ggml-cpu-impl.h"
|
||||
#include "ggml-impl.h"
|
||||
|
||||
|
||||
@ -1,8 +0,0 @@
|
||||
#pragma once
|
||||
|
||||
#include "ggml-cpu-traits.h"
|
||||
#include "ggml.h"
|
||||
|
||||
// GGML internal header
|
||||
|
||||
ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
|
||||
@ -506,3 +506,25 @@ void ggml_barrier(struct ggml_threadpool * tp);
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
|
||||
#define GGML_DO_PRAGMA_(x) _Pragma (#x)
|
||||
#define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
|
||||
#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__)
|
||||
// Note for Apple targets:
|
||||
// - clang: aliases are not supported on darwin
|
||||
// - all native kernels need to be implemented in both x86 and arm files
|
||||
// - on iOS, tvOS, and visionOS, if cmake cannot determine the target architecture, all `_generic` names are replaced by defines
|
||||
# define GGML_WEAK_ALIAS(name, alias)
|
||||
#elif defined(__GNUC__)
|
||||
// GCC/Clang on *nix
|
||||
# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(weak name = alias) // NOLINT
|
||||
#elif defined(_MSC_VER) && defined (_WIN64)
|
||||
// MSVC
|
||||
// Note: C name mangling varies across different calling conventions
|
||||
// see https://learn.microsoft.com/en-us/cpp/build/reference/decorated-names?view=msvc-170
|
||||
# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:" #name "=" #alias))
|
||||
#else
|
||||
# error "Unsupported compiler for GGML_WEAK_ALIAS"
|
||||
#endif
|
||||
|
||||
#define GGML_CPU_NATIVE_IMPL(name) GGML_WEAK_ALIAS(name, name ## _generic)
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@ -1,63 +0,0 @@
|
||||
#pragma once
|
||||
|
||||
#define GGML_COMMON_DECL_C
|
||||
#include "ggml-common.h"
|
||||
|
||||
#include "ggml.h"
|
||||
|
||||
// GGML CPU internal header
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
// Quantization
|
||||
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
|
||||
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
|
||||
void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
|
||||
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
|
||||
// Dot product
|
||||
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
@ -3,11 +3,11 @@
|
||||
|
||||
#include "ggml-backend-impl.h"
|
||||
#include "ggml-backend.h"
|
||||
#include "ggml-cpu-traits.h"
|
||||
#include "traits.h"
|
||||
#include "ggml-cpu-impl.h"
|
||||
#include "ggml-cpu.h"
|
||||
#include "ggml-impl.h"
|
||||
#include "ggml-cpu-quants.h"
|
||||
#include "quants.h"
|
||||
#include "ggml-threading.h"
|
||||
#include "unary-ops.h"
|
||||
#include "binary-ops.h"
|
||||
|
||||
@ -1,8 +1,8 @@
|
||||
#include "ggml-backend.h"
|
||||
#include "ggml-backend-impl.h"
|
||||
#include "ggml-cpu.h"
|
||||
#include "ggml-cpu-aarch64.h"
|
||||
#include "ggml-cpu-traits.h"
|
||||
#include "repack.h"
|
||||
#include "traits.h"
|
||||
#include "ggml-impl.h"
|
||||
#include "amx/amx.h"
|
||||
|
||||
@ -11,7 +11,7 @@
|
||||
#include <vector>
|
||||
|
||||
#ifdef GGML_USE_CPU_HBM
|
||||
# include "ggml-cpu-hbm.h"
|
||||
# include "hbm.h"
|
||||
#endif
|
||||
|
||||
#ifdef GGML_USE_CPU_KLEIDIAI
|
||||
@ -51,9 +51,9 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
|
||||
}
|
||||
#endif
|
||||
|
||||
#ifdef GGML_USE_CPU_AARCH64
|
||||
if (ggml_backend_cpu_aarch64_buffer_type()) {
|
||||
bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
|
||||
#ifdef GGML_USE_CPU_REPACK
|
||||
if (ggml_backend_cpu_repack_buffer_type()) {
|
||||
bufts.push_back(ggml_backend_cpu_repack_buffer_type());
|
||||
}
|
||||
#endif
|
||||
|
||||
@ -596,8 +596,8 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
|
||||
#ifdef GGML_USE_CPU_KLEIDIAI
|
||||
features.push_back({ "KLEIDIAI", "1" });
|
||||
#endif
|
||||
#ifdef GGML_USE_CPU_AARCH64
|
||||
features.push_back({ "AARCH64_REPACK", "1" });
|
||||
#ifdef GGML_USE_CPU_REPACK
|
||||
features.push_back({ "REPACK", "1" });
|
||||
#endif
|
||||
|
||||
features.push_back({ nullptr, nullptr });
|
||||
|
||||
@ -5,7 +5,7 @@
|
||||
#include "ggml-cpu.h"
|
||||
#include "ggml-impl.h"
|
||||
|
||||
#include "ggml-cpu-hbm.h"
|
||||
#include "hbm.h"
|
||||
|
||||
// buffer type HBM
|
||||
|
||||
@ -26,7 +26,7 @@
|
||||
#include "ggml-impl.h"
|
||||
#include "ggml-backend-impl.h"
|
||||
#include "ggml-threading.h"
|
||||
#include "ggml-cpu-traits.h"
|
||||
#include "traits.h"
|
||||
|
||||
#include "kernels.h"
|
||||
|
||||
|
||||
1179
ggml/src/ggml-cpu/quants.c
Normal file
1179
ggml/src/ggml-cpu/quants.c
Normal file
File diff suppressed because it is too large
Load Diff
116
ggml/src/ggml-cpu/quants.h
Normal file
116
ggml/src/ggml-cpu/quants.h
Normal file
@ -0,0 +1,116 @@
|
||||
#pragma once
|
||||
|
||||
#define GGML_COMMON_DECL_C
|
||||
#include "ggml-common.h"
|
||||
|
||||
#include "ggml.h"
|
||||
|
||||
// GGML CPU internal header
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
// Quantization
|
||||
void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
|
||||
void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
|
||||
void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
|
||||
void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
|
||||
// Dot product
|
||||
void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
// Generic implementation
|
||||
void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
|
||||
void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
|
||||
void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
|
||||
void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
|
||||
|
||||
#if defined(GGML_CPU_GENERIC)
|
||||
#define quantize_row_q8_0_generic quantize_row_q8_0
|
||||
#define quantize_row_q8_1_generic quantize_row_q8_1
|
||||
#define quantize_row_q8_K_generic quantize_row_q8_K
|
||||
#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
|
||||
#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
|
||||
#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
|
||||
#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
|
||||
#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
|
||||
#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
|
||||
#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
|
||||
#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
|
||||
#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
|
||||
#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
|
||||
#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
|
||||
#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
|
||||
#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
|
||||
#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
|
||||
#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
|
||||
#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
|
||||
#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
|
||||
#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
|
||||
#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
|
||||
#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
|
||||
#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
|
||||
#endif
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
1566
ggml/src/ggml-cpu/repack.cpp
Normal file
1566
ggml/src/ggml-cpu/repack.cpp
Normal file
File diff suppressed because it is too large
Load Diff
119
ggml/src/ggml-cpu/repack.h
Normal file
119
ggml/src/ggml-cpu/repack.h
Normal file
@ -0,0 +1,119 @@
|
||||
#pragma once
|
||||
|
||||
#define GGML_COMMON_DECL_CPP
|
||||
#include "ggml-common.h"
|
||||
|
||||
#include "traits.h"
|
||||
#include "ggml.h"
|
||||
|
||||
// GGML internal header
|
||||
|
||||
ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
|
||||
|
||||
template <int K> constexpr int QK_0() {
|
||||
if constexpr (K == 4) {
|
||||
return QK4_0;
|
||||
}
|
||||
if constexpr (K == 8) {
|
||||
return QK8_0;
|
||||
}
|
||||
return -1;
|
||||
}
|
||||
|
||||
template <int K, int N> struct block {
|
||||
ggml_half d[N]; // deltas for N qK_0 blocks
|
||||
int8_t qs[(QK_0<K>() * N * K) / 8]; // quants for N qK_0 blocks
|
||||
};
|
||||
|
||||
// control size
|
||||
static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
|
||||
static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
|
||||
static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
|
||||
static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
|
||||
|
||||
using block_q4_0x4 = block<4, 4>;
|
||||
using block_q4_0x8 = block<4, 8>;
|
||||
using block_q8_0x4 = block<8, 4>;
|
||||
using block_q8_0x8 = block<8, 8>;
|
||||
|
||||
struct block_q4_Kx8 {
|
||||
ggml_half d[8]; // super-block scale for quantized scales
|
||||
ggml_half dmin[8]; // super-block scale for quantized mins
|
||||
uint8_t scales[96]; // scales and mins, quantized with 6 bits
|
||||
uint8_t qs[1024]; // 4--bit quants
|
||||
};
|
||||
|
||||
static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
|
||||
|
||||
struct block_q8_Kx4 {
|
||||
float d[4]; // delta
|
||||
int8_t qs[QK_K * 4]; // quants
|
||||
int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
|
||||
};
|
||||
|
||||
static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
|
||||
|
||||
struct block_iq4_nlx4 {
|
||||
ggml_half d[4]; // deltas for 4 iq4_nl blocks
|
||||
uint8_t qs[QK4_NL * 2]; // nibbles / quants for 4 iq4_nl blocks
|
||||
};
|
||||
|
||||
static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
|
||||
|
||||
#if defined(__cplusplus)
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
// Workaround for clang:
|
||||
// clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous''
|
||||
// repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call)
|
||||
#if defined(GGML_CPU_CLANG_WORKAROUND) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
|
||||
void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
|
||||
void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
|
||||
void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
|
||||
void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
#endif // !defined(__clang__)
|
||||
|
||||
// Native implementations
|
||||
void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
|
||||
void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
|
||||
void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
|
||||
void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
|
||||
|
||||
#if defined(GGML_CPU_GENERIC)
|
||||
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
|
||||
#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
|
||||
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
|
||||
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
|
||||
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
|
||||
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
|
||||
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
|
||||
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
|
||||
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
|
||||
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
|
||||
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
|
||||
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
|
||||
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
|
||||
#endif
|
||||
|
||||
#if defined(__cplusplus)
|
||||
} // extern "C"
|
||||
#endif
|
||||
@ -1,4 +1,4 @@
|
||||
#include "ggml-cpu-traits.h"
|
||||
#include "traits.h"
|
||||
|
||||
#include "ggml-backend-impl.h"
|
||||
#include "ggml-backend.h"
|
||||
@ -466,9 +466,6 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
|
||||
#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
|
||||
}
|
||||
|
||||
// TODO: move to ggml-common.h
|
||||
static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
|
||||
|
||||
typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
|
||||
|
||||
static __device__ __forceinline__ float get_alibi_slope(
|
||||
|
||||
@ -2425,8 +2425,6 @@ void dequantize_row_iq1_m(const block_iq1_m * GGML_RESTRICT x, float * GGML_REST
|
||||
}
|
||||
}
|
||||
|
||||
static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
|
||||
|
||||
void dequantize_row_iq4_nl(const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
|
||||
assert(k % QK4_NL == 0);
|
||||
const int64_t nb = k / QK4_NL;
|
||||
|
||||
@ -149,8 +149,6 @@ typedef sycl::float2 dfloat2;
|
||||
|
||||
#define MMVQ_MAX_BATCH_SIZE 8
|
||||
|
||||
static const int8_t kvalues_iq4nl[16]={-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
|
||||
|
||||
static int g_all_sycl_device_count = -1;
|
||||
static bool g_ggml_backend_sycl_buffer_type_initialized = false;
|
||||
|
||||
|
||||
Reference in New Issue
Block a user