local-llm-rn 1.0.0

package/cpp/ggml/src/ggml-hexagon/htp/cmake-toolchain.cmake

@@ -0,0 +1,157 @@
+if (HEXAGON_TOOLCHAIN_INCLUDED)
+  return()
+endif()
+set(HEXAGON_TOOLCHAIN_INCLUDED true)
+
+#Cross Compiling for Hexagon
+set(HEXAGON TRUE)
+set(CMAKE_SYSTEM_NAME QURT)
+set(CMAKE_SYSTEM_PROCESSOR Hexagon)
+set(CMAKE_SYSTEM_VERSION "1") #${HEXAGON_PLATFORM_LEVEL})
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
+set(CUSTOM_RUNELF_PATH "")
+
+#To fix backward compatibility with EAI addon.
+if (NOT HEXAGON_SDK_ROOT)
+    set(HEXAGON_SDK_ROOT $ENV{HEXAGON_SDK_ROOT})
+endif()
+
+if (NOT HEXAGON_TOOLS_ROOT)
+    if (DEFINED ENV{HEXAGON_TOOLS_ROOT})
+        set(HEXAGON_TOOLS_ROOT $ENV{HEXAGON_TOOLS_ROOT})
+    endif()
+    if(NOT HEXAGON_TOOLS_ROOT)
+        set(HEXAGON_TOOLS_ROOT $ENV{DEFAULT_HEXAGON_TOOLS_ROOT})
+    endif()
+endif()
+
+file(TO_CMAKE_PATH "${HEXAGON_TOOLS_ROOT}" HEXAGON_TOOLS_ROOT)
+file(TO_CMAKE_PATH "${HEXAGON_SDK_ROOT}"   HEXAGON_SDK_ROOT)
+
+#Get the Binary extension of the Hexagon Toolchain
+if(CMAKE_HOST_SYSTEM_NAME STREQUAL Windows)
+    set(HEXAGON_TOOLCHAIN_SUFFIX .exe)
+endif()
+message(DEBUG "CMAKE_HOST_SYSTEM_NAME:${CMAKE_HOST_SYSTEM_NAME}")
+
+include(${HEXAGON_SDK_ROOT}/build/cmake/hexagon_arch.cmake)
+
+set(HEXAGON_TOOLCHAIN ${HEXAGON_TOOLS_ROOT})
+set(HEXAGON_LIB_DIR "${HEXAGON_TOOLCHAIN}/Tools/target/hexagon/lib")
+set(HEXAGON_ISS_DIR ${HEXAGON_TOOLCHAIN}/Tools/lib/iss)
+
+set(CMAKE_TRY_COMPILE_PLATFORM_VARIABLES
+    HEXAGON_SDK_ROOT
+    HEXAGON_TOOLS_ROOT
+)
+
+#QURT Related includes and linker flags
+set(V_ARCH ${HEXAGON_ARCH})
+set(_QURT_INSTALL_DIR "${HEXAGON_SDK_ROOT}/rtos/qurt/ADSP${V_ARCH}MP${V_ARCH_EXTN}")
+set(_QURT_INSTALL_DIR "${HEXAGON_SDK_ROOT}/rtos/qurt/compute${V_ARCH}${V_ARCH_EXTN}")
+
+if( ${TREE} MATCHES PAKMAN )
+    set(_QURT_INSTALL_DIR "${QURT_IMAGE_DIR}/compute${V_ARCH}${V_ARCH_EXTN}")
+endif()
+message(DEBUG "_QURT_INSTALL_DIR:${_QURT_INSTALL_DIR}")
+set(RTOS_DIR ${_QURT_INSTALL_DIR})
+set(QCC_DIR "${HEXAGON_QCC_DIR}/${V_ARCH}/G0")
+set(TARGET_DIR "${HEXAGON_LIB_DIR}/${V_ARCH}/G0")
+
+include_directories(
+    ${_QURT_INSTALL_DIR}/include
+    ${_QURT_INSTALL_DIR}/include/qurt
+    ${_QURT_INSTALL_DIR}/include/posix
+    )
+
+set(QURT_START_LINK_LIBS)
+set(QURT_START_LINK_LIBS
+    "${TARGET_DIR}/init.o"
+    "${RTOS_DIR}/lib/crt1.o"
+    "${RTOS_DIR}/lib/debugmon.o"
+    "${RTOS_DIR}/lib/libqurt.a"
+    "${TARGET_DIR}/libc.a"
+    "${TARGET_DIR}/libqcc.a"
+    "${TARGET_DIR}/libhexagon.a"
+    "${RTOS_DIR}/lib/libqurtcfs.a"
+    "${RTOS_DIR}/lib/libtimer_island.a"
+    "${RTOS_DIR}/lib/libtimer_main.a"
+    "${RTOS_DIR}/lib/libposix.a"
+    )
+STRING(REPLACE ";" " " QURT_START_LINK_LIBS "${QURT_START_LINK_LIBS}")
+
+set(QURT_END_LINK_LIBS
+    ${TARGET_DIR}/fini.o
+    )
+
+#Non QURT related includes and linker flags
+
+set(TARGET_DIR_NOOS "${HEXAGON_TOOLCHAIN}/Tools/target/hexagon/lib/${HEXAGON_ARCH}")
+
+if (NOT NO_WRAP_MEM_API)
+    set(WRAP_MALLOC   -Wl,--wrap=malloc)
+    set(WRAP_CALLOC   -Wl,--wrap=calloc)
+    set(WRAP_FREE     -Wl,--wrap=free)
+    set(WRAP_REALLOC  -Wl,--wrap=realloc)
+    set(WRAP_MEMALIGN -Wl,--wrap=memalign)
+endif()
+
+set(PIC_SHARED_LD_FLAGS
+    -mcpu=${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH}
+    -G0
+    -fpic
+    -Wl,-Bsymbolic
+    -Wl,-L${TARGET_DIR_NOOS}/G0/pic
+    -Wl,-L${HEXAGON_TOOLCHAIN}/Tools/target/hexagon/lib/
+    -Wl,--no-threads ${WRAP_MALLOC} ${WRAP_CALLOC} ${WRAP_FREE} ${WRAP_REALLOC} ${WRAP_MEMALIGN}
+    -shared
+    "-o <TARGET> <SONAME_FLAG><TARGET_SONAME>"
+    "<LINK_FLAGS>"
+    -Wl,--start-group
+    "<OBJECTS>"
+    "<LINK_LIBRARIES>"
+    -Wl,--end-group
+    -lc
+    )
+STRING(REPLACE ";" " " PIC_SHARED_LD_FLAGS "${PIC_SHARED_LD_FLAGS}")
+
+set(HEXAGON_PIC_SHARED_LINK_OPTIONS "${PIC_SHARED_LD_FLAGS}")
+
+#System include paths
+include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/incs)
+include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/incs/stddef)
+include_directories(SYSTEM ${HEXAGON_SDK_ROOT}/ipc/fastrpc/incs)
+
+#LLVM toolchain setup
+#Compiler paths, options and architecture
+set(CMAKE_C_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang${HEXAGON_TOOLCHAIN_SUFFIX})
+set(CMAKE_CXX_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang++${HEXAGON_TOOLCHAIN_SUFFIX})
+set(CMAKE_AR ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-ar${HEXAGON_TOOLCHAIN_SUFFIX})
+set(CMAKE_ASM_COMPILER ${HEXAGON_TOOLCHAIN}/Tools/bin/hexagon-clang++${HEXAGON_TOOLCHAIN_SUFFIX})
+set(HEXAGON_LINKER ${CMAKE_C_COMPILER})
+set(CMAKE_PREFIX_PATH ${HEXAGON_TOOLCHAIN}/Tools/target/hexagon)
+
+set(CMAKE_SHARED_LIBRARY_SONAME_C_FLAG   "-Wl,-soname,")
+set(CMAKE_SHARED_LIBRARY_SONAME_CXX_FLAG "-Wl,-soname,")
+
+#Compiler Options
+set(COMMON_FLAGS "-mcpu=hexagon${V_ARCH} -m${V_ARCH} -mhvx=${V_ARCH} -fvectorize -Wall -Werror -fno-zero-initialized-in-bss -G0 -fdata-sections -fpic ${XQF_ARGS}")
+
+set(CMAKE_CXX_FLAGS_DEBUG          "${COMMON_FLAGS} -O0 -D_DEBUG -g")
+set(CMAKE_CXX_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS} -O3 -g")
+set(CMAKE_CXX_FLAGS_RELEASE        "${COMMON_FLAGS} -O3")
+
+set(CMAKE_C_FLAGS_DEBUG            "${COMMON_FLAGS} -O0 -D_DEBUG -g")
+set(CMAKE_C_FLAGS_RELWITHDEBINFO   "${COMMON_FLAGS} -O3 -g")
+set(CMAKE_C_FLAGS_RELEASE          "${COMMON_FLAGS} -O3")
+
+set(CMAKE_ASM_FLAGS_DEBUG          "${COMMON_FLAGS} ${CMAKE_CXX_FLAGS_DEBUG}")
+set(CMAKE_ASM_FLAGS_RELEASE        "${COMMON_FLAGS} ${CMAKE_CXX_FLAGS_RELEASE}")
+set(CMAKE_ASM_FLAGS_RELWITHDEBINFO "${COMMON_FLAGS} ${CMAKE_CXX_FLAGS_RELWITHDEBINFO}" )
+
+#Linker Options
+set(CMAKE_C_CREATE_SHARED_LIBRARY   "${HEXAGON_LINKER} ${HEXAGON_PIC_SHARED_LINK_OPTIONS}")
+set(CMAKE_CXX_CREATE_SHARED_LIBRARY "${HEXAGON_LINKER} ${HEXAGON_PIC_SHARED_LINK_OPTIONS}")

package/cpp/ggml/src/ggml-hexagon/htp/cpy-ops.c

@@ -0,0 +1,251 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+#pragma clang diagnostic ignored "-Wunused-function"
+#pragma clang diagnostic ignored "-Wunused-but-set-variable"
+
+#include <HAP_farf.h>
+#include <HAP_perf.h>
+
+#include <math.h>
+#include <string.h>
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+#include "htp-ctx.h"
+#include "htp-msg.h"
+#include "htp-ops.h"
+#include "hvx-utils.h"
+
+struct htp_copy_context {
+    struct htp_ops_context * octx;
+
+    uint32_t          src0_type_size;
+    uint32_t          src0_block_size;
+
+    uint32_t          dst_type_size;
+    uint32_t          dst_block_size;
+
+    uint32_t          src0_blocks_per_row;
+    uint32_t          dst_blocks_per_row;
+
+    uint32_t          src0_nrows_per_thread;
+
+    void (*copy)(struct htp_copy_context * ct, struct htp_ops_context * octx, int nth, int ith);
+};
+
+#define cpy_preamble                       \
+    struct htp_tensor *src0 = &octx->src0; \
+    struct htp_tensor *dst  = &octx->dst;  \
+                                           \
+    const uint32_t ne00 = src0->ne[0];     \
+    const uint32_t ne01 = src0->ne[1];     \
+    const uint32_t ne02 = src0->ne[2];     \
+    const uint32_t ne03 = src0->ne[3];     \
+                                           \
+    const uint32_t nb00 = src0->nb[0];     \
+    const uint32_t nb01 = src0->nb[1];     \
+    const uint32_t nb02 = src0->nb[2];     \
+    const uint32_t nb03 = src0->nb[3];     \
+                                           \
+    const uint32_t  ne0 = dst->ne[0];      \
+    const uint32_t  ne1 = dst->ne[1];      \
+    const uint32_t  ne2 = dst->ne[2];      \
+    const uint32_t  ne3 = dst->ne[3];      \
+                                           \
+    const uint32_t  nb0 = dst->nb[0];      \
+    const uint32_t  nb1 = dst->nb[1];      \
+    const uint32_t  nb2 = dst->nb[2];      \
+    const uint32_t  nb3 = dst->nb[3];      \
+                                           \
+    const uint32_t   nr = ne01;
+
+static void cpy_thread_sametype_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
+    cpy_preamble;
+
+    // parallelize by src0 rows
+    const uint32_t dr  = ct->src0_nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
+
+    // copy by rows
+    for (uint32_t i03 = 0; i03 < ne03; i03++) {
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {
+            #pragma unroll(2)
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {
+                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
+                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                hex_l2fetch(src0_ptr, ne00 * ct->src0_type_size, nb01, 2);
+                hvx_copy_uu(dst_ptr, src0_ptr, ne00, ct->src0_type_size);
+            }
+        }
+    }
+}
+
+static void cpy_thread_sametype_reshape(struct htp_copy_context * ct, struct htp_ops_context * octx, int nth, int ith) {
+    cpy_preamble;
+
+    // parallelize by src0 rows
+    const uint32_t dr  = ct->src0_nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
+
+    // dst counters
+    int64_t k10 = 0;
+    int64_t i11 = 0;
+    int64_t i12 = 0;
+    int64_t i13 = 0;
+
+    // number of blocks in a row
+    const int64_t nk00 = ct->src0_blocks_per_row;
+    const int64_t nk0  = ct->dst_blocks_per_row;
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            k10 += nk00 * ir0;
+            while (k10 >= nk0) {
+                k10 -= nk0;
+                if (++i11 == ne1) {
+                    i11 = 0;
+                    if (++i12 == ne2) {
+                        i12 = 0;
+                        if (++i13 == ne3) {
+                            i13 = 0;
+                        }
+                    }
+                }
+            }
+            for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                for (int64_t k00 = 0; k00 < nk00; k00++) {
+                    const char * src0_ptr = ((char *) src0->data + k00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          char * dst_ptr  = ((char *)  dst->data + k10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+                    memcpy(dst_ptr, src0_ptr, ct->dst_type_size);
+
+                    if (++k10 == nk0) {
+                        k10 = 0;
+                        if (++i11 == ne1) {
+                            i11 = 0;
+                            if (++i12 == ne2) {
+                                i12 = 0;
+                                if (++i13 == ne3) {
+                                    i13 = 0;
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            k10 += nk00 * (ne01 - ir1);
+            while (k10 >= nk0) {
+                k10 -= nk0;
+                if (++i11 == ne1) {
+                    i11 = 0;
+                    if (++i12 == ne2) {
+                        i12 = 0;
+                        if (++i13 == ne3) {
+                            i13 = 0;
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void cpy_thread_f16_f32_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
+    cpy_preamble;
+
+    // parallelize by src0 rows
+    const uint32_t dr  = ct->src0_nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
+
+    // copy by rows
+    for (uint32_t i03 = 0; i03 < ne03; i03++) {
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {
+            #pragma unroll(2)
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {
+                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
+                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                hex_l2fetch(src0_ptr, ne00 * sizeof(float), nb01, 2);
+                hvx_copy_f16_f32_uu(dst_ptr, src0_ptr, ne00);
+            }
+        }
+    }
+}
+
+static void cpy_thread_f32_f16_sameshape(struct htp_copy_context * ct, struct htp_ops_context * octx, const int nth, const int ith) {
+    cpy_preamble;
+
+    // parallelize by src0 rows
+    const uint32_t dr  = ct->src0_nrows_per_thread;
+    const uint32_t ir0 = dr * ith;
+    const uint32_t ir1 = (ir0 + dr) < nr ? (ir0 + dr) : nr;
+
+    // copy by rows
+    for (uint32_t i03 = 0; i03 < ne03; i03++) {
+        for (uint32_t i02 = 0; i02 < ne02; i02++) {
+            #pragma unroll(2)
+            for (uint32_t i01 = ir0; i01 < ir1; i01++) {
+                uint8_t* dst_ptr  = (uint8_t*) dst->data  + i01*nb1  + i02*nb2  + i03*nb3;
+                uint8_t* src0_ptr = (uint8_t*) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                hex_l2fetch(src0_ptr, ne00 * sizeof(__fp16), nb01, 2);
+                hvx_copy_f32_f16_uu(dst_ptr, src0_ptr, ne00);
+            }
+        }
+    }
+}
+
+static void cpy_work_func(unsigned int n, unsigned int i, void *data) {
+    struct htp_copy_context *ct = (struct htp_copy_context *) data;
+    ct->copy(ct, ct->octx, n, i);
+}
+
+int op_cpy(struct htp_ops_context * octx) {
+    cpy_preamble;
+
+    struct htp_copy_context ct;
+    ct.octx = octx;
+
+    switch (src0->type) {
+    case HTP_TYPE_F32: ct.src0_type_size = 4; ct.src0_block_size = 1; ct.src0_blocks_per_row = ne00 / 1; break;
+    case HTP_TYPE_F16: ct.src0_type_size = 2; ct.src0_block_size = 1; ct.src0_blocks_per_row = ne00 / 1; break;
+    default:
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    switch (dst->type) {
+    case HTP_TYPE_F32: ct.dst_type_size = 4; ct.dst_block_size = 1; ct.dst_blocks_per_row = ne0 / 1; break;
+    case HTP_TYPE_F16: ct.dst_type_size = 2; ct.dst_block_size = 1; ct.dst_blocks_per_row = ne0 / 1; break;
+    default:
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    if (octx->flags & HTP_OPFLAGS_SKIP_COMPUTE) {
+        return HTP_STATUS_OK;
+    }
+
+    const bool sametype   = (src0->type == dst->type);
+    const bool transposed = (nb00 > nb01) || (nb0 > nb1);
+    const bool sameshape  = !transposed && (ne00 == ne0 && ne01 == ne1 && ne02 == ne2 && ne03 == ne3);
+
+    const uint32_t n_jobs = MIN(nr, octx->n_threads);
+    ct.src0_nrows_per_thread = (nr + n_jobs - 1) / n_jobs;
+
+    if (sametype && sameshape) {
+        ct.copy = cpy_thread_sametype_sameshape;
+    } else if (sameshape) {
+        /**/ if (dst->type == HTP_TYPE_F16 && src0->type == HTP_TYPE_F32)
+            ct.copy = cpy_thread_f16_f32_sameshape;
+        else if (dst->type == HTP_TYPE_F32 && src0->type == HTP_TYPE_F16)
+            ct.copy = cpy_thread_f32_f16_sameshape;
+        else
+            return HTP_STATUS_NO_SUPPORT;
+    } else if (sametype) {
+        ct.copy = cpy_thread_sametype_reshape;
+    } else {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    worker_pool_run_func(octx->ctx->worker_pool, cpy_work_func, &ct, n_jobs);
+
+    return HTP_STATUS_OK;
+}