From: Georgi Gerganov Date: Sat, 26 Oct 2024 06:44:39 +0000 (+0300) Subject: ggml : remove sync artifacts X-Git-Tag: upstream/0.0.1642~263 X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=162e232411ee98ceb0cccfa84886118d917d2123;p=pkg%2Fggml%2Fsources%2Fggml ggml : remove sync artifacts ggml-ci --- diff --git a/ggml/include/ggml-amx.h b/ggml/include/ggml-amx.h deleted file mode 100644 index 22b3f70f..00000000 --- a/ggml/include/ggml-amx.h +++ /dev/null @@ -1,25 +0,0 @@ -#pragma once - -#include "ggml.h" -#include "ggml-backend.h" - - -#ifdef __cplusplus -extern "C" { -#endif - -// buffer_type API -GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void); - -GGML_API bool ggml_backend_is_amx(ggml_backend_t backend); - -// backend API -GGML_API ggml_backend_t ggml_backend_amx_init(void); - -GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads); - -GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void); - -#ifdef __cplusplus -} -#endif diff --git a/ggml/include/ggml-cann.h b/ggml/include/ggml-cann.h deleted file mode 100644 index ca73211f..00000000 --- a/ggml/include/ggml-cann.h +++ /dev/null @@ -1,125 +0,0 @@ -/* - * Copyright (c) 2023-2024 The ggml authors - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in - * all copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING - * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS - * IN THE SOFTWARE. - */ - -#pragma once - -#include "ggml-backend.h" -#include "ggml.h" - -#ifdef __cplusplus -extern "C" { -#endif - -/** - * @brief Maximum number of CANN devices supported. - */ -#define GGML_CANN_MAX_DEVICES 16 - -/** - * @brief Initializes the CANN backend for a specified device. - * - * This function initializes the CANN backend for the given device. - * It verifies the device index, allocates a context, and creates a backend - * instance. - * - * @param device The index of the device to initialize. - * @return A pointer to the initialized backend instance, or nullptr on failure. - */ -GGML_API GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device); - -/** - * @brief Checks if a given backend is a CANN backend. - * - * This function verifies if the provided backend is a CANN backend by comparing - * its GUID with the CANN backend's GUID. - * - * @param backend The backend instance to check. - * @return True if the backend is a CANN backend, false otherwise. - */ -GGML_API GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend); - -/** - * @brief Retrieves the CANN buffer type for a specified device. - * - * This function initializes and returns the buffer type interface associated - * with the given device. It ensures thread-safe access using a mutex. - * - * @param device The device index for which to retrieve the buffer type. - * @return A pointer to the buffer type interface for the specified device, or - * nullptr if the device index is out of range. - */ -GGML_API GGML_CALL ggml_backend_buffer_type_t -ggml_backend_cann_buffer_type(int32_t device); - -/** - * @brief Retrieves the number of CANN devices available. - * - * This function returns the number of CANN devices available based on - * information obtained from `ggml_cann_info()`. - * - * @return The number of CANN devices available. - */ -GGML_API GGML_CALL int32_t ggml_backend_cann_get_device_count(void); - -/** - * @brief Retrieves the description of a specific CANN device. - * - * This function sets the specified device, retrieves the SoC name, - * and writes it into the provided description buffer. - * - * @param device The device index to retrieve the description for. - * @param description Pointer to a buffer where the description will be written. - * @param description_size Size of the description buffer. - */ -GGML_API GGML_CALL void ggml_backend_cann_get_device_description( - int32_t device, char* description, size_t description_size); - -/** - * @brief Retrieves the memory information of a specific CANN device. - * - * This function sets the specified device, retrieves the free and total - * memory information of the specified type (ACL_HBM_MEM), and stores them - * in the provided pointers. - * - * @param device The device index to retrieve memory information for. - * @param free Pointer to a variable where the free memory size will be stored. - * @param total Pointer to a variable where the total memory size will be - * stored. - */ -GGML_API GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device, - size_t* free, - size_t* total); - -/** - * @brief Set the logging callback for GGML. - * - * This function sets the logging callback and user data for logging. - * - * @param log_callback The logging callback to set. - * @param user_data User data to pass to the logging callback. - */ -GGML_API void ggml_backend_cann_log_set_callback(ggml_log_callback log_callback, - void* user_data); - -#ifdef __cplusplus -} -#endif diff --git a/ggml/src/ggml-aarch64.c b/ggml/src/ggml-aarch64.c deleted file mode 100644 index af53dea1..00000000 --- a/ggml/src/ggml-aarch64.c +++ /dev/null @@ -1,2193 +0,0 @@ -// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd. -#define GGML_COMMON_IMPL_C -#include "ggml-common.h" - -#include "ggml-quants.h" -#include "ggml-impl.h" - -#include -#include -#include -#include -#include // for qsort -#include // for GGML_ASSERT - -#include "ggml-aarch64.h" - -#if defined(__GNUC__) -#pragma GCC diagnostic ignored "-Woverlength-strings" -#endif - -#define UNUSED GGML_UNUSED - -// Functions to create the interleaved data layout formats - -// interleave 4 block_q4_0s in blocks of blck_size_interleave -// returns an interleaved block_q4_0x4 -// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks -// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave -// -// - in : an array of block_q4_0 pointers -// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of -// blck_size_interleave bytes -// - xor_mask : the mask to convert the nibbles in block_q4_0 quants bytes -// from bias offset form to pure sign form (this saves subtract -// operations durin unpacking) -// -static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) { - block_q4_0x4 out; - - for (int i = 0; i < 4; i++) { - out.d[i] = in[i].d; - } - - for (int i = 0; i < QK4_0 * 2; i++) { - int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (i % blck_size_interleave); - - out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask; - } - - return out; -} - -// interleave 8 block_q4_0s in blocks of blck_size_interleave -// returns an interleaved block_q4_0x8 -// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks -// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave -static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) { - block_q4_0x8 out; - - for (int i = 0; i < 8; i++) { - out.d[i] = in[i].d; - } - - for (int i = 0; i < QK4_0 * 4; i++) { - int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave; - int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave; - src_offset += (i % blck_size_interleave); - - out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask; - } - - return out; -} - -void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) { - assert(QK8_0 == 32); - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; - - block_q8_0x4 * restrict y = (block_q8_0x4 *) vy; - -#if defined(__ARM_NEON) - float32x4_t srcv[4][8]; - float id[4]; - - for (int i = 0; i < nb; i++) { - float32x4_t asrcv[8]; - float32x4_t amaxv[8]; - - for (int row_iter = 0; row_iter < 4; row_iter++) { - for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j); - for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]); - - for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]); - for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]); - for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]); - - const float amax = vmaxvq_f32(amaxv[0]); - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); - } - - for (int j = 0; j < 8; j++) { - float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]); - int32x4_t vi = vcvtnq_s32_f32(v); - y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3); - - v = vmulq_n_f32(srcv[1][j], id[1]); - vi = vcvtnq_s32_f32(v); - y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0); - y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1); - y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2); - y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3); - - v = vmulq_n_f32(srcv[2][j], id[2]); - vi = vcvtnq_s32_f32(v); - y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0); - y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1); - y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2); - y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3); - - v = vmulq_n_f32(srcv[3][j], id[3]); - vi = vcvtnq_s32_f32(v); - y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0); - y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1); - y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2); - y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3); - } - } -#else - // scalar - const int blck_size_interleave = 4; - float srcv[4][QK8_0]; - float id[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; - amax = MAX(amax, fabsf(srcv[row_iter][j])); - } - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); - } - - for (int j = 0; j < QK8_0 * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - - float x0 = srcv[src_id][src_offset] * id[src_id]; - y[i].qs[j] = roundf(x0); - } - } -#endif -} - -void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) { - assert(QK8_0 == 32); - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; - - block_q8_0x4 * restrict y = (block_q8_0x4 *) vy; - -#if defined(__ARM_NEON) - float32x4_t srcv[4][8]; - float id[4]; - - for (int i = 0; i < nb; i++) { - float32x4_t asrcv[8]; - float32x4_t amaxv[8]; - - for (int row_iter = 0; row_iter < 4; row_iter++) { - for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j); - for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]); - - for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]); - for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]); - for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]); - - const float amax = vmaxvq_f32(amaxv[0]); - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); - } - - for (int j = 0; j < 4; j++) { - float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]); - int32x4_t vi = vcvtnq_s32_f32(v); - y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3); - v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]); - vi = vcvtnq_s32_f32(v); - y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0); - y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1); - y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2); - y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3); - - v = vmulq_n_f32(srcv[1][2 * j], id[1]); - vi = vcvtnq_s32_f32(v); - y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0); - y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1); - y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2); - y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3); - v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]); - vi = vcvtnq_s32_f32(v); - y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0); - y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1); - y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2); - y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3); - - v = vmulq_n_f32(srcv[2][2 * j], id[2]); - vi = vcvtnq_s32_f32(v); - y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0); - y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1); - y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2); - y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3); - v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]); - vi = vcvtnq_s32_f32(v); - y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0); - y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1); - y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2); - y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3); - - v = vmulq_n_f32(srcv[3][2 * j], id[3]); - vi = vcvtnq_s32_f32(v); - y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0); - y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1); - y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2); - y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3); - v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]); - vi = vcvtnq_s32_f32(v); - y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0); - y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1); - y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2); - y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3); - } - } -#else - // scalar - const int blck_size_interleave = 8; - float srcv[4][QK8_0]; - float id[4]; - - for (int i = 0; i < nb; i++) { - for (int row_iter = 0; row_iter < 4; row_iter++) { - float amax = 0.0f; // absolute max - - for (int j = 0; j < QK8_0; j++) { - srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j]; - amax = MAX(amax, fabsf(srcv[row_iter][j])); - } - - const float d = amax / ((1 << 7) - 1); - id[row_iter] = d ? 1.0f / d : 0.0f; - - y[i].d[row_iter] = GGML_FP32_TO_FP16(d); - } - - for (int j = 0; j < QK8_0 * 4; j++) { - int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; - int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; - src_offset += (j % blck_size_interleave); - - float x0 = srcv[src_id][src_offset] * id[src_id]; - y[i].qs[j] = roundf(x0); - } - } -#endif -} - -void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) { - assert(nrow == 4); - UNUSED(nrow); - if (blck_size_interleave == 4) { - quantize_q8_0_4x4(x, vy, n_per_row); - } else if (blck_size_interleave == 8) { - quantize_q8_0_4x8(x, vy, n_per_row); - } else { - assert(false); - } -} - -static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) { - assert(n_per_row % QK4_0 == 0); - const int nb = n_per_row / QK4_0; - - void * out_ptr = NULL; - if (nrows_interleaved == 8) { - out_ptr = (block_q4_0x8 *) dst; - } - else if (nrows_interleaved == 4) { - out_ptr = (block_q4_0x4 *) dst; - } - assert(nrows_interleaved <= 8); - block_q4_0 dst_tmp[8]; - - for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) { - - for (int64_t x = 0; x < nb; x++) { - - for (int i = 0; i < nrows_interleaved; i++ ) { - quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0); - } - - if (nrows_interleaved == 8) { - *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88); - out_ptr = (block_q4_0x8 *) out_ptr + 1; - } - else if (nrows_interleaved == 4) { - *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88); - out_ptr = (block_q4_0x4 *) out_ptr + 1; - } - } - } - - return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0)); -} - -size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - if (!quant_weights) { - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4); - } - else { - assert(false); - return 0; - } -} - -size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - if (!quant_weights) { - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8); - } - else { - assert(false); - return 0; - } -} - -size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) { - if (!quant_weights) { - return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8); - } - else { - assert(false); - return 0; - } -} - -void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { - const int qk = QK8_0; - const int nb = n / qk; - const int ncols_interleaved = 4; - const int blocklen = 4; - - 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(__ARM_FEATURE_SVE) - if (svcntw() == 8) { - GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) && - "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) && - "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance"); -#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - - __asm__ __volatile__( - "movi v31.16b, #0x4\n" - "movi v30.16b, #0xf0\n" - "add %x[b_ptr], %x[b_ptr], #0x8\n" - "1:" // Column loop - "add x22, %x[a_ptr], #0x2\n" - "movi v29.16b, #0x0\n" - "mov x21, %x[nb]\n" - "2:" // Block loop - "ldr q28, [%x[b_ptr], #0x0]\n" - "ldr q27, [x22, #0x0]\n" - "movi v26.4s, #0x0\n" - "sub x20, x22, #0x2\n" - "ldr q25, [x22, #0x10]\n" - "ldr q24, [%x[b_ptr], #0x10]\n" - "sub x21, x21, #0x1\n" - "add x22, x22, #0x22\n" - "ldr q23, [%x[b_ptr], #0x20]\n" - "ldr q22, [%x[b_ptr], #0x30]\n" - "ld1r { v21.8h }, [x20]\n" - "ldr q20, [%x[b_ptr], #-0x8]\n" - "sshl v16.16b, v28.16b, v31.16b\n" - "and v28.16b, v28.16b, v30.16b\n" - "sshl v19.16b, v24.16b, v31.16b\n" - "and v24.16b, v24.16b, v30.16b\n" - "add %x[b_ptr], %x[b_ptr], #0x48\n" - "sshl v18.16b, v23.16b, v31.16b\n" - "and v23.16b, v23.16b, v30.16b\n" - ".inst 0x4f9be21a // sdot v26.4s, v16.16b, v27.4b[0]\n" - "sshl v17.16b, v22.16b, v31.16b\n" - "and v22.16b, v22.16b, v30.16b\n" - "fcvtl v21.4s, v21.4h\n" - "fcvtl v16.4s, v20.4h\n" - ".inst 0x4f99e39a // sdot v26.4s, v28.16b, v25.4b[0]\n" - "fmul v16.4s, v16.4s, v21.4s\n" - ".inst 0x4fbbe27a // sdot v26.4s, v19.16b, v27.4b[1]\n" - ".inst 0x4fb9e31a // sdot v26.4s, v24.16b, v25.4b[1]\n" - ".inst 0x4f9bea5a // sdot v26.4s, v18.16b, v27.4b[2]\n" - ".inst 0x4f99eafa // sdot v26.4s, v23.16b, v25.4b[2]\n" - ".inst 0x4fbbea3a // sdot v26.4s, v17.16b, v27.4b[3]\n" - ".inst 0x4fb9eada // sdot v26.4s, v22.16b, v25.4b[3]\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "fmla v29.4s, v26.4s, v16.4s\n" - "cbnz x21, 2b\n" - "sub %x[nc], %x[nc], #0x4\n" - "str q29, [%x[res_ptr], #0x0]\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "cbnz %x[nc], 1b\n" - : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) - : [a_ptr] "r" (a_ptr), [nb] "r" (nb) - : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22" - ); -#else - float sumf[4]; - 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_0x4 * b_ptr = (const block_q4_0x4 *) 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]; - } -#endif -} - -void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { - const int qk = QK8_0; - const int nb = n / qk; - const int ncols_interleaved = 4; - 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(__ARM_FEATURE_SVE) - if (svcntw() == 8) { - GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) && - "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - - __asm__ __volatile__( - "movi v2.16b, #0x4\n" - "movi v1.16b, #0xf0\n" - "add %x[b_ptr], %x[b_ptr], #0x8\n" - "1:" // Column loop - "add x23, %x[a_ptr], #0x2\n" - "movi v0.16b, #0x0\n" - "mov x22, %x[nb]\n" - "2:" // Block loop - "ldr q31, [%x[b_ptr], #0x0]\n" - "ldr q30, [%x[b_ptr], #0x10]\n" - "mov x21, x23\n" - "movi v29.4s, #0x0\n" - "ldr q28, [%x[b_ptr], #0x20]\n" - "ldr q27, [%x[b_ptr], #0x30]\n" - "movi v26.4s, #0x0\n" - "sub x20, x23, #0x2\n" - "ld1r { v25.8h }, [x20]\n" - "ldr q24, [%x[b_ptr], #-0x8]\n" - "sub x22, x22, #0x1\n" - "add x23, x23, #0x22\n" - "ld1r { v23.2d }, [x21], #0x8\n" - "sshl v22.16b, v31.16b, v2.16b\n" - "sshl v16.16b, v30.16b, v2.16b\n" - "add %x[b_ptr], %x[b_ptr], #0x48\n" - "ld1r { v21.2d }, [x21], #0x8\n" - "sshl v20.16b, v28.16b, v2.16b\n" - "sshl v19.16b, v27.16b, v2.16b\n" - "ld1r { v18.2d }, [x21], #0x8\n" - "ld1r { v17.2d }, [x21], #0x8\n" - "and v31.16b, v31.16b, v1.16b\n" - "and v30.16b, v30.16b, v1.16b\n" - ".inst 0x4e9796dd // sdot v29.4s, v22.16b, v23.16b\n" - ".inst 0x4e97961a // sdot v26.4s, v16.16b, v23.16b\n" - "and v28.16b, v28.16b, v1.16b\n" - "and v27.16b, v27.16b, v1.16b\n" - "fcvtl v25.4s, v25.4h\n" - "fcvtl v16.4s, v24.4h\n" - ".inst 0x4e95969d // sdot v29.4s, v20.16b, v21.16b\n" - ".inst 0x4e95967a // sdot v26.4s, v19.16b, v21.16b\n" - "fmul v16.4s, v16.4s, v25.4s\n" - ".inst 0x4e9297fd // sdot v29.4s, v31.16b, v18.16b\n" - ".inst 0x4e9297da // sdot v26.4s, v30.16b, v18.16b\n" - ".inst 0x4e91979d // sdot v29.4s, v28.16b, v17.16b\n" - ".inst 0x4e91977a // sdot v26.4s, v27.16b, v17.16b\n" - "addp v29.4s, v29.4s, v26.4s\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "fmla v0.4s, v29.4s, v16.4s\n" - "cbnz x22, 2b\n" - "sub %x[nc], %x[nc], #0x4\n" - "str q0, [%x[res_ptr], #0x0]\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "cbnz %x[nc], 1b\n" - : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) - : [a_ptr] "r" (a_ptr), [nb] "r" (nb) - : "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23" - ); -#elif defined(__ARM_NEON) && defined(__aarch64__) - GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " - "performance"); -#else - float sumf[4]; - 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_0x4 * b_ptr = (const block_q4_0x4 *) 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]; - } -#endif -} - -void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * 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(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - if (svcntw() == 8) { - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - - __asm__ __volatile__( - "ptrue p0.b\n" - "add %x[b_ptr], %x[b_ptr], #0x10\n" - "1:" // Column loop - "add x22, %x[a_ptr], #0x2\n" - "mov z31.b, #0x0\n" - "mov x21, %x[nb]\n" - "2:" // Block loop - "ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n" - "ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n" - "mov z28.s, #0x0\n" - "mov z27.s, #0x0\n" - "ld1rd { z26.d }, p0/Z, [x22]\n" - "ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n" - "sub x20, x22, #0x2\n" - "sub x21, x21, #0x1\n" - "ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n" - "ld1rd { z23.d }, p0/Z, [x22, #8]\n" - "lsl z22.b, z30.b, #0x4\n" - "lsl z16.b, z29.b, #0x4\n" - "and z30.b, z30.b, #0xf0\n" - "and z29.b, z29.b, #0xf0\n" - "ld1rd { z21.d }, p0/Z, [x22, #16]\n" - "ld1rd { z20.d }, p0/Z, [x22, #24]\n" - "lsl z19.b, z25.b, #0x4\n" - "and z25.b, z25.b, #0xf0\n" - "ld1rh { z17.h }, p0/Z, [x20]\n" - "ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n" - "sdot z28.s, z22.b, z26.b\n" - "sdot z27.s, z16.b, z26.b\n" - "lsl z16.b, z24.b, #0x4\n" - "add x22, x22, #0x22\n" - "and z24.b, z24.b, #0xf0\n" - "add %x[b_ptr], %x[b_ptr], #0x90\n" - "fcvt z17.s, p0/m, z17.h\n" - "fcvt z18.s, p0/m, z18.h\n" - "sdot z28.s, z19.b, z23.b\n" - "sdot z27.s, z16.b, z23.b\n" - "fmul z18.s, z18.s, z17.s\n" - "sdot z28.s, z30.b, z21.b\n" - "sdot z27.s, z29.b, z21.b\n" - "sdot z28.s, z25.b, z20.b\n" - "sdot z27.s, z24.b, z20.b\n" - "uzp1 z17.s, z28.s, z27.s\n" - "uzp2 z16.s, z28.s, z27.s\n" - "add z17.s, z17.s, z16.s\n" - "asr z17.s, z17.s, #0x4\n" - "scvtf z17.s, p0/m, z17.s\n" - "fmla z31.s, p0/M, z17.s, z18.s\n" - "cbnz x21, 2b\n" - "sub %x[nc], %x[nc], #0x8\n" - "st1w { z31.s }, p0, [%x[res_ptr]]\n" - "add %x[res_ptr], %x[res_ptr], #0x20\n" - "cbnz %x[nc], 1b\n" - : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc) - : [a_ptr] "r" (a_ptr), [nb] "r" (nb) - : "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" - ); - return; - } - else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { - GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) && - "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal " - "performance"); - } - else if (ggml_cpu_has_neon()) { - GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 " - "quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - GGML_ASSERT(ggml_cpu_has_sve() && - "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance"); -#elif defined(__ARM_NEON) && defined(__aarch64__) - GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " - "performance"); -#else - 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]; - } -#endif -} - -void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { - const int qk = QK8_0; - const int nb = n / qk; - const int ncols_interleaved = 4; - const int blocklen = 4; - - 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(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) - if (svcntw() == 8) { - GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) && - "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) && - "__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance"); -#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - size_t res_stride = bs * sizeof(float); - - __asm__ __volatile__( - "mov x10, %x[nr]\n" - "mov x9, #0x88\n" - "cmp x10, #0x10\n" - "mul x9, %x[nb], x9\n" - "blt 4f\n" - "1:" // Row loop - "add x28, %x[b_ptr], #0x8\n" - "mov x27, %x[nc]\n" - "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" - "2:" // Column loop - "add x25, %x[a_ptr], #0x8\n" - "movi v15.16b, #0x0\n" - "movi v19.16b, #0x0\n" - "mov x24, %x[nb]\n" - "add x23, x25, x9\n" - "movi v18.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "add x22, x23, x9\n" - "movi v11.16b, #0x0\n" - "movi v13.16b, #0x0\n" - "add x21, x22, x9\n" - "movi v23.16b, #0x0\n" - "movi v16.16b, #0x0\n" - "movi v25.16b, #0x0\n" - "movi v7.16b, #0x0\n" - "movi v0.16b, #0x0\n" - "movi v4.16b, #0x0\n" - "movi v5.16b, #0x0\n" - "movi v21.16b, #0x0\n" - "movi v8.16b, #0x0\n" - "movi v1.16b, #0x0\n" - "3:" // Block loop - "ldr q3, [x28, #0x0]\n" - "ldr q31, [x25, #0x0]\n" - "movi v28.16b, #0x4\n" - "movi v10.4s, #0x0\n" - "ldr q22, [x28, #0x10]\n" - "ldr q6, [x25, #0x10]\n" - "movi v29.4s, #0x0\n" - "movi v9.4s, #0x0\n" - "ldr q27, [x28, #0x20]\n" - "ldr q30, [x28, #0x30]\n" - "movi v20.4s, #0x0\n" - "movi v24.16b, #0xf0\n" - "ldr d2, [x25, #-0x8]\n" - "ldr d26, [x23, #-0x8]\n" - "sshl v12.16b, v3.16b, v28.16b\n" - "sub x20, x28, #0x8\n" - "ldr d17, [x20, #0x0]\n" - "and v3.16b, v3.16b, v24.16b\n" - "subs x24, x24, #0x1\n" - "add x28, x28, #0x48\n" - ".inst 0x4f9fe18a // sdot v10.4s, v12.16b, v31.4b[0]\n" - ".inst 0x4fbfe19d // sdot v29.4s, v12.16b, v31.4b[1]\n" - ".inst 0x4f9fe989 // sdot v9.4s, v12.16b, v31.4b[2]\n" - ".inst 0x4fbfe994 // sdot v20.4s, v12.16b, v31.4b[3]\n" - "sshl v31.16b, v22.16b, v28.16b\n" - "and v22.16b, v22.16b, v24.16b\n" - "fcvtl v17.4s, v17.4h\n" - "fcvtl v2.4s, v2.4h\n" - "fcvtl v26.4s, v26.4h\n" - ".inst 0x4f86e3ea // sdot v10.4s, v31.16b, v6.4b[0]\n" - ".inst 0x4fa6e3fd // sdot v29.4s, v31.16b, v6.4b[1]\n" - ".inst 0x4f86ebe9 // sdot v9.4s, v31.16b, v6.4b[2]\n" - ".inst 0x4fa6ebf4 // sdot v20.4s, v31.16b, v6.4b[3]\n" - "sshl v6.16b, v27.16b, v28.16b\n" - "sshl v28.16b, v30.16b, v28.16b\n" - "and v27.16b, v27.16b, v24.16b\n" - "and v30.16b, v30.16b, v24.16b\n" - "ldr q24, [x25, #0x20]\n" - ".inst 0x4f98e0ca // sdot v10.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8c9 // sdot v9.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8d4 // sdot v20.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x30]\n" - ".inst 0x4f98e38a // sdot v10.4s, v28.16b, v24.4b[0]\n" - ".inst 0x4fb8e39d // sdot v29.4s, v28.16b, v24.4b[1]\n" - ".inst 0x4f98eb89 // sdot v9.4s, v28.16b, v24.4b[2]\n" - ".inst 0x4fb8eb94 // sdot v20.4s, v28.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x40]\n" - ".inst 0x4f98e06a // sdot v10.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e869 // sdot v9.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e874 // sdot v20.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x50]\n" - ".inst 0x4f98e2ca // sdot v10.4s, v22.16b, v24.4b[0]\n" - ".inst 0x4fb8e2dd // sdot v29.4s, v22.16b, v24.4b[1]\n" - ".inst 0x4f98eac9 // sdot v9.4s, v22.16b, v24.4b[2]\n" - ".inst 0x4fb8ead4 // sdot v20.4s, v22.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x60]\n" - ".inst 0x4f98e36a // sdot v10.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb69 // sdot v9.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb74 // sdot v20.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x25, #0x70]\n" - "add x25, x25, #0x88\n" - ".inst 0x4f98e3ca // sdot v10.4s, v30.16b, v24.4b[0]\n" - ".inst 0x4fb8e3dd // sdot v29.4s, v30.16b, v24.4b[1]\n" - ".inst 0x4f98ebc9 // sdot v9.4s, v30.16b, v24.4b[2]\n" - ".inst 0x4fb8ebd4 // sdot v20.4s, v30.16b, v24.4b[3]\n" - "fmul v24.4s, v17.4s, v2.s[0]\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v15.4s, v10.4s, v24.4s\n" - "ldr q24, [x23, #0x0]\n" - "fmul v10.4s, v17.4s, v2.s[1]\n" - "fmla v19.4s, v29.4s, v10.4s\n" - "ldr q10, [x23, #0x10]\n" - "fmul v29.4s, v17.4s, v2.s[2]\n" - "fmul v2.4s, v17.4s, v2.s[3]\n" - "fmla v18.4s, v9.4s, v29.4s\n" - "movi v9.4s, #0x0\n" - "movi v29.4s, #0x0\n" - ".inst 0x4f98e189 // sdot v9.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e19d // sdot v29.4s, v12.16b, v24.4b[1]\n" - "fmla v14.4s, v20.4s, v2.4s\n" - "movi v20.4s, #0x0\n" - "movi v2.4s, #0x0\n" - ".inst 0x4f98e994 // sdot v20.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x20]\n" - ".inst 0x4f8ae3e9 // sdot v9.4s, v31.16b, v10.4b[0]\n" - ".inst 0x4faae3fd // sdot v29.4s, v31.16b, v10.4b[1]\n" - ".inst 0x4f8aebf4 // sdot v20.4s, v31.16b, v10.4b[2]\n" - ".inst 0x4faaebe2 // sdot v2.4s, v31.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x30]\n" - ".inst 0x4f98e0c9 // sdot v9.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8d4 // sdot v20.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x40]\n" - ".inst 0x4f8ae389 // sdot v9.4s, v28.16b, v10.4b[0]\n" - ".inst 0x4faae39d // sdot v29.4s, v28.16b, v10.4b[1]\n" - ".inst 0x4f8aeb94 // sdot v20.4s, v28.16b, v10.4b[2]\n" - ".inst 0x4faaeb82 // sdot v2.4s, v28.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x50]\n" - ".inst 0x4f98e069 // sdot v9.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e874 // sdot v20.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x23, #0x60]\n" - ".inst 0x4f8ae2c9 // sdot v9.4s, v22.16b, v10.4b[0]\n" - ".inst 0x4faae2dd // sdot v29.4s, v22.16b, v10.4b[1]\n" - ".inst 0x4f8aead4 // sdot v20.4s, v22.16b, v10.4b[2]\n" - ".inst 0x4faaeac2 // sdot v2.4s, v22.16b, v10.4b[3]\n" - "ldr q10, [x23, #0x70]\n" - "add x23, x23, #0x88\n" - ".inst 0x4f98e369 // sdot v9.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb74 // sdot v20.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x0]\n" - ".inst 0x4f8ae3c9 // sdot v9.4s, v30.16b, v10.4b[0]\n" - ".inst 0x4faae3dd // sdot v29.4s, v30.16b, v10.4b[1]\n" - ".inst 0x4f8aebd4 // sdot v20.4s, v30.16b, v10.4b[2]\n" - ".inst 0x4faaebc2 // sdot v2.4s, v30.16b, v10.4b[3]\n" - "fmul v10.4s, v17.4s, v26.s[0]\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "fmla v11.4s, v9.4s, v10.4s\n" - "ldr q9, [x22, #0x10]\n" - "fmul v10.4s, v17.4s, v26.s[1]\n" - "fmla v13.4s, v29.4s, v10.4s\n" - "ldr d29, [x22, #-0x8]\n" - "fmul v10.4s, v17.4s, v26.s[2]\n" - "fmul v26.4s, v17.4s, v26.s[3]\n" - "fcvtl v29.4s, v29.4h\n" - "fmla v23.4s, v20.4s, v10.4s\n" - "movi v20.4s, #0x0\n" - "movi v10.4s, #0x0\n" - "fmla v16.4s, v2.4s, v26.4s\n" - "movi v26.4s, #0x0\n" - "movi v2.4s, #0x0\n" - ".inst 0x4f98e194 // sdot v20.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" - ".inst 0x4f98e99a // sdot v26.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x20]\n" - ".inst 0x4f89e3f4 // sdot v20.4s, v31.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" - ".inst 0x4f89ebfa // sdot v26.4s, v31.16b, v9.4b[2]\n" - ".inst 0x4fa9ebe2 // sdot v2.4s, v31.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x30]\n" - ".inst 0x4f98e0d4 // sdot v20.4s, v6.16b, v24.4b[0]\n" - ".inst 0x4fb8e0ca // sdot v10.4s, v6.16b, v24.4b[1]\n" - ".inst 0x4f98e8da // sdot v26.4s, v6.16b, v24.4b[2]\n" - ".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x40]\n" - ".inst 0x4f89e394 // sdot v20.4s, v28.16b, v9.4b[0]\n" - ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" - ".inst 0x4f89eb9a // sdot v26.4s, v28.16b, v9.4b[2]\n" - ".inst 0x4fa9eb82 // sdot v2.4s, v28.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x50]\n" - ".inst 0x4f98e074 // sdot v20.4s, v3.16b, v24.4b[0]\n" - ".inst 0x4fb8e06a // sdot v10.4s, v3.16b, v24.4b[1]\n" - ".inst 0x4f98e87a // sdot v26.4s, v3.16b, v24.4b[2]\n" - ".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n" - "ldr q24, [x22, #0x60]\n" - ".inst 0x4f89e2d4 // sdot v20.4s, v22.16b, v9.4b[0]\n" - ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" - ".inst 0x4f89eada // sdot v26.4s, v22.16b, v9.4b[2]\n" - ".inst 0x4fa9eac2 // sdot v2.4s, v22.16b, v9.4b[3]\n" - "ldr q9, [x22, #0x70]\n" - "add x22, x22, #0x88\n" - ".inst 0x4f98e374 // sdot v20.4s, v27.16b, v24.4b[0]\n" - ".inst 0x4fb8e36a // sdot v10.4s, v27.16b, v24.4b[1]\n" - ".inst 0x4f98eb7a // sdot v26.4s, v27.16b, v24.4b[2]\n" - ".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n" - "ldr q24, [x21, #0x0]\n" - ".inst 0x4f89e3d4 // sdot v20.4s, v30.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ca // sdot v10.4s, v30.16b, v9.4b[1]\n" - ".inst 0x4f89ebda // sdot v26.4s, v30.16b, v9.4b[2]\n" - ".inst 0x4fa9ebc2 // sdot v2.4s, v30.16b, v9.4b[3]\n" - "fmul v9.4s, v17.4s, v29.s[0]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "fmla v25.4s, v20.4s, v9.4s\n" - "ldr q9, [x21, #0x10]\n" - "fmul v20.4s, v17.4s, v29.s[1]\n" - "fmla v7.4s, v10.4s, v20.4s\n" - "ldr d20, [x21, #-0x8]\n" - "fmul v10.4s, v17.4s, v29.s[2]\n" - "fmul v29.4s, v17.4s, v29.s[3]\n" - "fcvtl v20.4s, v20.4h\n" - "fmla v0.4s, v26.4s, v10.4s\n" - "movi v26.4s, #0x0\n" - "movi v10.4s, #0x0\n" - "fmla v4.4s, v2.4s, v29.4s\n" - "movi v2.4s, #0x0\n" - "movi v29.4s, #0x0\n" - ".inst 0x4f98e19a // sdot v26.4s, v12.16b, v24.4b[0]\n" - ".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n" - ".inst 0x4f98e982 // sdot v2.4s, v12.16b, v24.4b[2]\n" - ".inst 0x4fb8e99d // sdot v29.4s, v12.16b, v24.4b[3]\n" - "ldr q12, [x21, #0x20]\n" - "fmul v24.4s, v17.4s, v20.s[0]\n" - ".inst 0x4f89e3fa // sdot v26.4s, v31.16b, v9.4b[0]\n" - ".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n" - ".inst 0x4f89ebe2 // sdot v2.4s, v31.16b, v9.4b[2]\n" - ".inst 0x4fa9ebfd // sdot v29.4s, v31.16b, v9.4b[3]\n" - "ldr q9, [x21, #0x30]\n" - "fmul v31.4s, v17.4s, v20.s[1]\n" - ".inst 0x4f8ce0da // sdot v26.4s, v6.16b, v12.4b[0]\n" - ".inst 0x4face0ca // sdot v10.4s, v6.16b, v12.4b[1]\n" - ".inst 0x4f8ce8c2 // sdot v2.4s, v6.16b, v12.4b[2]\n" - ".inst 0x4face8dd // sdot v29.4s, v6.16b, v12.4b[3]\n" - "ldr q12, [x21, #0x40]\n" - "fmul v6.4s, v17.4s, v20.s[2]\n" - "fmul v20.4s, v17.4s, v20.s[3]\n" - ".inst 0x4f89e39a // sdot v26.4s, v28.16b, v9.4b[0]\n" - ".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n" - ".inst 0x4f89eb82 // sdot v2.4s, v28.16b, v9.4b[2]\n" - ".inst 0x4fa9eb9d // sdot v29.4s, v28.16b, v9.4b[3]\n" - "ldr q9, [x21, #0x50]\n" - ".inst 0x4f8ce07a // sdot v26.4s, v3.16b, v12.4b[0]\n" - ".inst 0x4face06a // sdot v10.4s, v3.16b, v12.4b[1]\n" - ".inst 0x4f8ce862 // sdot v2.4s, v3.16b, v12.4b[2]\n" - ".inst 0x4face87d // sdot v29.4s, v3.16b, v12.4b[3]\n" - "ldr q12, [x21, #0x60]\n" - ".inst 0x4f89e2da // sdot v26.4s, v22.16b, v9.4b[0]\n" - ".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n" - ".inst 0x4f89eac2 // sdot v2.4s, v22.16b, v9.4b[2]\n" - ".inst 0x4fa9eadd // sdot v29.4s, v22.16b, v9.4b[3]\n" - "ldr q17, [x21, #0x70]\n" - "add x21, x21, #0x88\n" - ".inst 0x4f8ce37a // sdot v26.4s, v27.16b, v12.4b[0]\n" - ".inst 0x4face36a // sdot v10.4s, v27.16b, v12.4b[1]\n" - ".inst 0x4f8ceb62 // sdot v2.4s, v27.16b, v12.4b[2]\n" - ".inst 0x4faceb7d // sdot v29.4s, v27.16b, v12.4b[3]\n" - ".inst 0x4f91e3da // sdot v26.4s, v30.16b, v17.4b[0]\n" - ".inst 0x4fb1e3ca // sdot v10.4s, v30.16b, v17.4b[1]\n" - ".inst 0x4f91ebc2 // sdot v2.4s, v30.16b, v17.4b[2]\n" - ".inst 0x4fb1ebdd // sdot v29.4s, v30.16b, v17.4b[3]\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "scvtf v10.4s, v10.4s, #0x4\n" - "fmla v5.4s, v26.4s, v24.4s\n" - "scvtf v2.4s, v2.4s, #0x4\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "fmla v21.4s, v10.4s, v31.4s\n" - "fmla v8.4s, v2.4s, v6.4s\n" - "fmla v1.4s, v29.4s, v20.4s\n" - "bgt 3b\n" - "mov x20, %x[res_ptr]\n" - "subs x27, x27, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "str q15, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q19, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q18, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q14, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q11, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q13, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q23, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q16, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q25, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q7, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q0, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q4, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q5, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q21, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q8, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q1, [x20, #0x0]\n" - "bne 2b\n" - "mov x20, #0x4\n" - "sub x10, x10, #0x10\n" - "cmp x10, #0x10\n" - "mov %x[res_ptr], x26\n" - "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" - "bge 1b\n" - "4:" // Row loop skip - "cbz x10, 9f\n" - "5:" // Row tail: Row loop - "add x24, %x[b_ptr], #0x8\n" - "mov x23, %x[nc]\n" - "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" - "6:" // Row tail: Column loop - "movi v15.16b, #0x0\n" - "movi v19.16b, #0x0\n" - "add x25, %x[a_ptr], #0x8\n" - "mov x21, %x[nb]\n" - "movi v18.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "7:" // Row tail: Block loop - "ldr q7, [x24, #0x0]\n" - "ldr q5, [x25, #0x0]\n" - "movi v9.16b, #0x4\n" - "movi v4.4s, #0x0\n" - "ldr q3, [x24, #0x10]\n" - "ldr q2, [x25, #0x10]\n" - "movi v1.4s, #0x0\n" - "movi v0.4s, #0x0\n" - "ldr q13, [x24, #0x20]\n" - "ldr q31, [x25, #0x20]\n" - "movi v30.4s, #0x0\n" - "movi v29.16b, #0xf0\n" - "ldr q28, [x24, #0x30]\n" - "ldr q27, [x25, #0x30]\n" - "sshl v20.16b, v7.16b, v9.16b\n" - "sub x20, x24, #0x8\n" - "ldr q26, [x25, #0x40]\n" - "ldr q25, [x25, #0x50]\n" - "sshl v17.16b, v3.16b, v9.16b\n" - "and v7.16b, v7.16b, v29.16b\n" - "ldr q24, [x25, #0x60]\n" - "ldr q16, [x25, #0x70]\n" - "sshl v22.16b, v13.16b, v9.16b\n" - "and v3.16b, v3.16b, v29.16b\n" - "ldr d21, [x20, #0x0]\n" - "ldr d12, [x25, #-0x8]\n" - ".inst 0x4f85e284 // sdot v4.4s, v20.16b, v5.4b[0]\n" - ".inst 0x4fa5e281 // sdot v1.4s, v20.16b, v5.4b[1]\n" - ".inst 0x4f85ea80 // sdot v0.4s, v20.16b, v5.4b[2]\n" - ".inst 0x4fa5ea9e // sdot v30.4s, v20.16b, v5.4b[3]\n" - "sshl v9.16b, v28.16b, v9.16b\n" - "subs x21, x21, #0x1\n" - "and v13.16b, v13.16b, v29.16b\n" - "and v28.16b, v28.16b, v29.16b\n" - "add x25, x25, #0x88\n" - "add x24, x24, #0x48\n" - "fcvtl v21.4s, v21.4h\n" - "fcvtl v12.4s, v12.4h\n" - ".inst 0x4f82e224 // sdot v4.4s, v17.16b, v2.4b[0]\n" - ".inst 0x4fa2e221 // sdot v1.4s, v17.16b, v2.4b[1]\n" - ".inst 0x4f82ea20 // sdot v0.4s, v17.16b, v2.4b[2]\n" - ".inst 0x4fa2ea3e // sdot v30.4s, v17.16b, v2.4b[3]\n" - "fmul v11.4s, v21.4s, v12.s[0]\n" - "fmul v23.4s, v21.4s, v12.s[1]\n" - "fmul v17.4s, v21.4s, v12.s[2]\n" - ".inst 0x4f9fe2c4 // sdot v4.4s, v22.16b, v31.4b[0]\n" - "fmul v6.4s, v21.4s, v12.s[3]\n" - ".inst 0x4fbfe2c1 // sdot v1.4s, v22.16b, v31.4b[1]\n" - ".inst 0x4f9feac0 // sdot v0.4s, v22.16b, v31.4b[2]\n" - ".inst 0x4fbfeade // sdot v30.4s, v22.16b, v31.4b[3]\n" - ".inst 0x4f9be124 // sdot v4.4s, v9.16b, v27.4b[0]\n" - ".inst 0x4fbbe121 // sdot v1.4s, v9.16b, v27.4b[1]\n" - ".inst 0x4f9be920 // sdot v0.4s, v9.16b, v27.4b[2]\n" - ".inst 0x4fbbe93e // sdot v30.4s, v9.16b, v27.4b[3]\n" - ".inst 0x4f9ae0e4 // sdot v4.4s, v7.16b, v26.4b[0]\n" - ".inst 0x4fbae0e1 // sdot v1.4s, v7.16b, v26.4b[1]\n" - ".inst 0x4f9ae8e0 // sdot v0.4s, v7.16b, v26.4b[2]\n" - ".inst 0x4fbae8fe // sdot v30.4s, v7.16b, v26.4b[3]\n" - ".inst 0x4f99e064 // sdot v4.4s, v3.16b, v25.4b[0]\n" - ".inst 0x4fb9e061 // sdot v1.4s, v3.16b, v25.4b[1]\n" - ".inst 0x4f99e860 // sdot v0.4s, v3.16b, v25.4b[2]\n" - ".inst 0x4fb9e87e // sdot v30.4s, v3.16b, v25.4b[3]\n" - ".inst 0x4f98e1a4 // sdot v4.4s, v13.16b, v24.4b[0]\n" - ".inst 0x4fb8e1a1 // sdot v1.4s, v13.16b, v24.4b[1]\n" - ".inst 0x4f98e9a0 // sdot v0.4s, v13.16b, v24.4b[2]\n" - ".inst 0x4fb8e9be // sdot v30.4s, v13.16b, v24.4b[3]\n" - ".inst 0x4f90e384 // sdot v4.4s, v28.16b, v16.4b[0]\n" - ".inst 0x4fb0e381 // sdot v1.4s, v28.16b, v16.4b[1]\n" - ".inst 0x4f90eb80 // sdot v0.4s, v28.16b, v16.4b[2]\n" - ".inst 0x4fb0eb9e // sdot v30.4s, v28.16b, v16.4b[3]\n" - "scvtf v4.4s, v4.4s, #0x4\n" - "scvtf v1.4s, v1.4s, #0x4\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "fmla v15.4s, v4.4s, v11.4s\n" - "scvtf v30.4s, v30.4s, #0x4\n" - "fmla v19.4s, v1.4s, v23.4s\n" - "fmla v18.4s, v0.4s, v17.4s\n" - "fmla v14.4s, v30.4s, v6.4s\n" - "bgt 7b\n" - "mov x20, %x[res_ptr]\n" - "cmp x10, #0x1\n" - "str q15, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x2\n" - "str q19, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x3\n" - "str q18, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "str q14, [x20, #0x0]\n" - "8:" // Row tail: Accumulator store skip - "subs x23, x23, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "bne 6b\n" - "subs x10, x10, #0x4\n" - "add %x[a_ptr], %x[a_ptr], x9\n" - "mov %x[res_ptr], x22\n" - "bgt 5b\n" - "9:" // Row tail: Row loop skip - : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) - : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" - ); -#else - float sumf[4][4]; - 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_0x4 * b_ptr = (const block_q4_0x4 *) 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]; - } - } - } -#endif -} - -void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) { - const int qk = QK8_0; - const int nb = n / qk; - const int ncols_interleaved = 4; - 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(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) - if (svcntw() == 8) { - GGML_ASSERT(!(ggml_cpu_has_sve() && (svcntw() == 8)) && - "__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - size_t res_stride = bs * sizeof(float); - - __asm__ __volatile__( - "mov x10, %x[nr]\n" - "mov x9, #0x88\n" - "cmp x10, #0x10\n" - "mul x9, %x[nb], x9\n" - "blt 4f\n" - "1:" // Row loop - "add x28, %x[b_ptr], #0x8\n" - "mov x27, %x[nc]\n" - "add x26, %x[res_ptr], %x[res_stride], LSL #4\n" - "2:" // Column loop - "add x25, %x[a_ptr], #0x8\n" - "movi v2.16b, #0x0\n" - "movi v10.16b, #0x0\n" - "mov x24, %x[nb]\n" - "add x23, x25, x9\n" - "movi v12.16b, #0x0\n" - "movi v28.16b, #0x0\n" - "add x22, x23, x9\n" - "movi v11.16b, #0x0\n" - "movi v13.16b, #0x0\n" - "add x21, x22, x9\n" - "movi v22.16b, #0x0\n" - "movi v23.16b, #0x0\n" - "movi v25.16b, #0x0\n" - "movi v5.16b, #0x0\n" - "movi v7.16b, #0x0\n" - "movi v4.16b, #0x0\n" - "movi v6.16b, #0x0\n" - "movi v30.16b, #0x0\n" - "movi v24.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "3:" // Block loop - "ldr q21, [x28, #0x0]\n" - "ldr q16, [x28, #0x10]\n" - "movi v1.16b, #0x4\n" - "movi v19.4s, #0x0\n" - "ldr q27, [x25, #0x0]\n" - "ldr q15, [x25, #0x10]\n" - "movi v26.4s, #0x0\n" - "movi v18.4s, #0x0\n" - "ldr q29, [x28, #0x20]\n" - "ldr q3, [x28, #0x30]\n" - "movi v17.4s, #0x0\n" - "movi v0.16b, #0xf0\n" - "ldr d20, [x25, #-0x8]\n" - "ldr d9, [x23, #-0x8]\n" - "sshl v8.16b, v21.16b, v1.16b\n" - "sshl v31.16b, v16.16b, v1.16b\n" - "and v21.16b, v21.16b, v0.16b\n" - "and v16.16b, v16.16b, v0.16b\n" - "sub x20, x28, #0x8\n" - "subs x24, x24, #0x1\n" - "add x28, x28, #0x48\n" - ".inst 0x4e88a773 // smmla v19.4s, v27.16b, v8.16b\n" - ".inst 0x4e9fa77a // smmla v26.4s, v27.16b, v31.16b\n" - "ldr q27, [x25, #0x20]\n" - ".inst 0x4e88a5f2 // smmla v18.4s, v15.16b, v8.16b\n" - ".inst 0x4e9fa5f1 // smmla v17.4s, v15.16b, v31.16b\n" - "sshl v15.16b, v29.16b, v1.16b\n" - "sshl v1.16b, v3.16b, v1.16b\n" - "and v29.16b, v29.16b, v0.16b\n" - "and v3.16b, v3.16b, v0.16b\n" - "ldr q0, [x25, #0x30]\n" - "fcvtl v20.4s, v20.4h\n" - ".inst 0x4e8fa773 // smmla v19.4s, v27.16b, v15.16b\n" - "fcvtl v9.4s, v9.4h\n" - ".inst 0x4e81a77a // smmla v26.4s, v27.16b, v1.16b\n" - "ldr q27, [x25, #0x40]\n" - ".inst 0x4e8fa412 // smmla v18.4s, v0.16b, v15.16b\n" - ".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n" - "ldr q0, [x25, #0x50]\n" - ".inst 0x4e95a773 // smmla v19.4s, v27.16b, v21.16b\n" - ".inst 0x4e90a77a // smmla v26.4s, v27.16b, v16.16b\n" - "ldr q27, [x25, #0x60]\n" - ".inst 0x4e95a412 // smmla v18.4s, v0.16b, v21.16b\n" - ".inst 0x4e90a411 // smmla v17.4s, v0.16b, v16.16b\n" - "ldr q0, [x25, #0x70]\n" - "add x25, x25, #0x88\n" - ".inst 0x4e9da773 // smmla v19.4s, v27.16b, v29.16b\n" - ".inst 0x4e83a77a // smmla v26.4s, v27.16b, v3.16b\n" - "ldr d27, [x20, #0x0]\n" - ".inst 0x4e9da412 // smmla v18.4s, v0.16b, v29.16b\n" - ".inst 0x4e83a411 // smmla v17.4s, v0.16b, v3.16b\n" - "fcvtl v27.4s, v27.4h\n" - "uzp1 v0.2d, v19.2d, v26.2d\n" - "uzp2 v26.2d, v19.2d, v26.2d\n" - "fmul v19.4s, v27.4s, v20.s[0]\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "scvtf v26.4s, v26.4s, #0x4\n" - "fmla v2.4s, v0.4s, v19.4s\n" - "ldr q19, [x23, #0x0]\n" - "uzp1 v0.2d, v18.2d, v17.2d\n" - "uzp2 v18.2d, v18.2d, v17.2d\n" - "fmul v17.4s, v27.4s, v20.s[1]\n" - "scvtf v0.4s, v0.4s, #0x4\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "fmla v10.4s, v26.4s, v17.4s\n" - "ldr q17, [x23, #0x10]\n" - "fmul v26.4s, v27.4s, v20.s[2]\n" - "fmul v20.4s, v27.4s, v20.s[3]\n" - "fmla v12.4s, v0.4s, v26.4s\n" - "ldr d0, [x22, #-0x8]\n" - "ldr d26, [x21, #-0x8]\n" - "fcvtl v0.4s, v0.4h\n" - "fmla v28.4s, v18.4s, v20.4s\n" - "movi v20.4s, #0x0\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" - ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" - "ldr q19, [x23, #0x20]\n" - "fcvtl v26.4s, v26.4h\n" - ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" - ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" - "ldr q19, [x23, #0x40]\n" - ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" - ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" - "ldr q19, [x23, #0x60]\n" - ".inst 0x4e9da674 // smmla v20.4s, v19.16b, v29.16b\n" - ".inst 0x4e83a672 // smmla v18.4s, v19.16b, v3.16b\n" - "uzp1 v19.2d, v20.2d, v18.2d\n" - "scvtf v19.4s, v19.4s, #0x4\n" - "uzp2 v20.2d, v20.2d, v18.2d\n" - "fmul v18.4s, v27.4s, v9.s[0]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v11.4s, v19.4s, v18.4s\n" - "ldr q18, [x22, #0x0]\n" - "fmul v19.4s, v27.4s, v9.s[1]\n" - "fmla v13.4s, v20.4s, v19.4s\n" - "movi v19.4s, #0x0\n" - "movi v20.4s, #0x0\n" - ".inst 0x4e88a633 // smmla v19.4s, v17.16b, v8.16b\n" - ".inst 0x4e9fa634 // smmla v20.4s, v17.16b, v31.16b\n" - "ldr q17, [x23, #0x30]\n" - ".inst 0x4e8fa633 // smmla v19.4s, v17.16b, v15.16b\n" - ".inst 0x4e81a634 // smmla v20.4s, v17.16b, v1.16b\n" - "ldr q17, [x23, #0x50]\n" - ".inst 0x4e95a633 // smmla v19.4s, v17.16b, v21.16b\n" - ".inst 0x4e90a634 // smmla v20.4s, v17.16b, v16.16b\n" - "ldr q17, [x23, #0x70]\n" - "add x23, x23, #0x88\n" - ".inst 0x4e9da633 // smmla v19.4s, v17.16b, v29.16b\n" - ".inst 0x4e83a634 // smmla v20.4s, v17.16b, v3.16b\n" - "uzp1 v17.2d, v19.2d, v20.2d\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "uzp2 v20.2d, v19.2d, v20.2d\n" - "fmul v19.4s, v27.4s, v9.s[2]\n" - "fmul v9.4s, v27.4s, v9.s[3]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v22.4s, v17.4s, v19.4s\n" - "ldr q17, [x22, #0x10]\n" - "movi v19.4s, #0x0\n" - ".inst 0x4e88a653 // smmla v19.4s, v18.16b, v8.16b\n" - "fmla v23.4s, v20.4s, v9.4s\n" - "movi v20.4s, #0x0\n" - "movi v9.4s, #0x0\n" - ".inst 0x4e9fa654 // smmla v20.4s, v18.16b, v31.16b\n" - "ldr q18, [x22, #0x20]\n" - ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" - ".inst 0x4e8fa653 // smmla v19.4s, v18.16b, v15.16b\n" - ".inst 0x4e81a654 // smmla v20.4s, v18.16b, v1.16b\n" - "ldr q18, [x22, #0x40]\n" - ".inst 0x4e95a653 // smmla v19.4s, v18.16b, v21.16b\n" - ".inst 0x4e90a654 // smmla v20.4s, v18.16b, v16.16b\n" - "ldr q18, [x22, #0x60]\n" - ".inst 0x4e9da653 // smmla v19.4s, v18.16b, v29.16b\n" - ".inst 0x4e83a654 // smmla v20.4s, v18.16b, v3.16b\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e9fa632 // smmla v18.4s, v17.16b, v31.16b\n" - "ldr q17, [x22, #0x30]\n" - ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" - ".inst 0x4e81a632 // smmla v18.4s, v17.16b, v1.16b\n" - "ldr q17, [x22, #0x50]\n" - ".inst 0x4e95a629 // smmla v9.4s, v17.16b, v21.16b\n" - ".inst 0x4e90a632 // smmla v18.4s, v17.16b, v16.16b\n" - "ldr q17, [x22, #0x70]\n" - "add x22, x22, #0x88\n" - ".inst 0x4e9da629 // smmla v9.4s, v17.16b, v29.16b\n" - ".inst 0x4e83a632 // smmla v18.4s, v17.16b, v3.16b\n" - "uzp1 v17.2d, v19.2d, v20.2d\n" - "uzp2 v20.2d, v19.2d, v20.2d\n" - "fmul v19.4s, v27.4s, v0.s[0]\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "fmla v25.4s, v17.4s, v19.4s\n" - "ldr q19, [x21, #0x0]\n" - "fmul v17.4s, v27.4s, v0.s[1]\n" - "fmla v5.4s, v20.4s, v17.4s\n" - "ldr q17, [x21, #0x10]\n" - "uzp1 v20.2d, v9.2d, v18.2d\n" - "uzp2 v9.2d, v9.2d, v18.2d\n" - "fmul v18.4s, v27.4s, v0.s[2]\n" - "fmul v0.4s, v27.4s, v0.s[3]\n" - "scvtf v20.4s, v20.4s, #0x4\n" - "scvtf v9.4s, v9.4s, #0x4\n" - "fmla v7.4s, v20.4s, v18.4s\n" - "movi v20.4s, #0x0\n" - "movi v18.4s, #0x0\n" - ".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n" - ".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n" - "ldr q19, [x21, #0x20]\n" - "fmla v4.4s, v9.4s, v0.4s\n" - "movi v9.4s, #0x0\n" - "movi v0.4s, #0x0\n" - ".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n" - "fmul v8.4s, v27.4s, v26.s[0]\n" - ".inst 0x4e9fa620 // smmla v0.4s, v17.16b, v31.16b\n" - "ldr q17, [x21, #0x30]\n" - ".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n" - "fmul v31.4s, v27.4s, v26.s[1]\n" - ".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n" - "ldr q19, [x21, #0x40]\n" - ".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n" - "fmul v15.4s, v27.4s, v26.s[2]\n" - "fmul v27.4s, v27.4s, v26.s[3]\n" - ".inst 0x4e81a620 // smmla v0.4s, v17.16b, v1.16b\n" - "ldr q1, [x21, #0x50]\n" - ".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n" - ".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n" - "ldr q26, [x21, #0x60]\n" - ".inst 0x4e95a429 // smmla v9.4s, v1.16b, v21.16b\n" - ".inst 0x4e90a420 // smmla v0.4s, v1.16b, v16.16b\n" - "ldr q21, [x21, #0x70]\n" - "add x21, x21, #0x88\n" - ".inst 0x4e9da754 // smmla v20.4s, v26.16b, v29.16b\n" - ".inst 0x4e83a752 // smmla v18.4s, v26.16b, v3.16b\n" - ".inst 0x4e9da6a9 // smmla v9.4s, v21.16b, v29.16b\n" - ".inst 0x4e83a6a0 // smmla v0.4s, v21.16b, v3.16b\n" - "uzp1 v29.2d, v20.2d, v18.2d\n" - "uzp2 v21.2d, v20.2d, v18.2d\n" - "scvtf v29.4s, v29.4s, #0x4\n" - "uzp1 v18.2d, v9.2d, v0.2d\n" - "uzp2 v16.2d, v9.2d, v0.2d\n" - "scvtf v21.4s, v21.4s, #0x4\n" - "fmla v6.4s, v29.4s, v8.4s\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "scvtf v16.4s, v16.4s, #0x4\n" - "fmla v30.4s, v21.4s, v31.4s\n" - "fmla v24.4s, v18.4s, v15.4s\n" - "fmla v14.4s, v16.4s, v27.4s\n" - "bgt 3b\n" - "mov x20, %x[res_ptr]\n" - "subs x27, x27, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "str q2, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q10, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q12, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q28, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q11, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q13, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q22, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q23, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q25, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q5, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q7, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q4, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q6, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q30, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q24, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "str q14, [x20, #0x0]\n" - "bne 2b\n" - "mov x20, #0x4\n" - "sub x10, x10, #0x10\n" - "cmp x10, #0x10\n" - "mov %x[res_ptr], x26\n" - "madd %x[a_ptr], x20, x9, %x[a_ptr]\n" - "bge 1b\n" - "4:" // Row loop skip - "cbz x10, 9f\n" - "5:" // Row tail: Row loop - "add x24, %x[b_ptr], #0x8\n" - "mov x23, %x[nc]\n" - "add x22, %x[res_ptr], %x[res_stride], LSL #2\n" - "6:" // Row tail: Column loop - "movi v2.16b, #0x0\n" - "movi v10.16b, #0x0\n" - "add x25, %x[a_ptr], #0x8\n" - "mov x21, %x[nb]\n" - "movi v12.16b, #0x0\n" - "movi v28.16b, #0x0\n" - "7:" // Row tail: Block loop - "ldr q6, [x24, #0x0]\n" - "ldr q5, [x24, #0x10]\n" - "movi v17.16b, #0x4\n" - "movi v8.4s, #0x0\n" - "ldr q4, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "movi v27.4s, #0x0\n" - "movi v0.4s, #0x0\n" - "ldr q31, [x24, #0x20]\n" - "ldr q14, [x24, #0x30]\n" - "movi v29.4s, #0x0\n" - "movi v22.16b, #0xf0\n" - "ldr q11, [x25, #0x20]\n" - "ldr q23, [x25, #0x30]\n" - "sshl v21.16b, v6.16b, v17.16b\n" - "sshl v16.16b, v5.16b, v17.16b\n" - "ldr q20, [x25, #0x40]\n" - "ldr q26, [x25, #0x50]\n" - "and v6.16b, v6.16b, v22.16b\n" - "and v5.16b, v5.16b, v22.16b\n" - "ldr q25, [x25, #0x60]\n" - "ldr q3, [x25, #0x70]\n" - "sshl v19.16b, v31.16b, v17.16b\n" - "sshl v18.16b, v14.16b, v17.16b\n" - "ldr d17, [x25, #-0x8]\n" - ".inst 0x4e95a488 // smmla v8.4s, v4.16b, v21.16b\n" - ".inst 0x4e90a49b // smmla v27.4s, v4.16b, v16.16b\n" - "and v31.16b, v31.16b, v22.16b\n" - ".inst 0x4e95a5a0 // smmla v0.4s, v13.16b, v21.16b\n" - ".inst 0x4e90a5bd // smmla v29.4s, v13.16b, v16.16b\n" - "and v14.16b, v14.16b, v22.16b\n" - "sub x20, x24, #0x8\n" - "ldr d16, [x20, #0x0]\n" - "subs x21, x21, #0x1\n" - "add x25, x25, #0x88\n" - "fcvtl v17.4s, v17.4h\n" - "add x24, x24, #0x48\n" - ".inst 0x4e93a568 // smmla v8.4s, v11.16b, v19.16b\n" - ".inst 0x4e92a57b // smmla v27.4s, v11.16b, v18.16b\n" - ".inst 0x4e93a6e0 // smmla v0.4s, v23.16b, v19.16b\n" - ".inst 0x4e92a6fd // smmla v29.4s, v23.16b, v18.16b\n" - "fcvtl v16.4s, v16.4h\n" - ".inst 0x4e86a688 // smmla v8.4s, v20.16b, v6.16b\n" - ".inst 0x4e85a69b // smmla v27.4s, v20.16b, v5.16b\n" - "fmul v23.4s, v16.4s, v17.s[0]\n" - "fmul v21.4s, v16.4s, v17.s[1]\n" - "fmul v1.4s, v16.4s, v17.s[2]\n" - "fmul v20.4s, v16.4s, v17.s[3]\n" - ".inst 0x4e86a740 // smmla v0.4s, v26.16b, v6.16b\n" - ".inst 0x4e85a75d // smmla v29.4s, v26.16b, v5.16b\n" - ".inst 0x4e9fa728 // smmla v8.4s, v25.16b, v31.16b\n" - ".inst 0x4e8ea73b // smmla v27.4s, v25.16b, v14.16b\n" - ".inst 0x4e9fa460 // smmla v0.4s, v3.16b, v31.16b\n" - ".inst 0x4e8ea47d // smmla v29.4s, v3.16b, v14.16b\n" - "uzp1 v19.2d, v8.2d, v27.2d\n" - "uzp2 v18.2d, v8.2d, v27.2d\n" - "scvtf v19.4s, v19.4s, #0x4\n" - "uzp1 v17.2d, v0.2d, v29.2d\n" - "uzp2 v16.2d, v0.2d, v29.2d\n" - "scvtf v18.4s, v18.4s, #0x4\n" - "fmla v2.4s, v19.4s, v23.4s\n" - "scvtf v17.4s, v17.4s, #0x4\n" - "scvtf v16.4s, v16.4s, #0x4\n" - "fmla v10.4s, v18.4s, v21.4s\n" - "fmla v12.4s, v17.4s, v1.4s\n" - "fmla v28.4s, v16.4s, v20.4s\n" - "bgt 7b\n" - "mov x20, %x[res_ptr]\n" - "cmp x10, #0x1\n" - "str q2, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x2\n" - "str q10, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x10, #0x3\n" - "str q12, [x20, #0x0]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "str q28, [x20, #0x0]\n" - "8:" // Row tail: Accumulator store skip - "subs x23, x23, #0x4\n" - "add %x[res_ptr], %x[res_ptr], #0x10\n" - "bne 6b\n" - "subs x10, x10, #0x4\n" - "add %x[a_ptr], %x[a_ptr], x9\n" - "mov %x[res_ptr], x22\n" - "bgt 5b\n" - "9:" // Row tail: Row loop skip - : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) - : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" - ); -#elif defined(__ARM_NEON) && defined(__aarch64__) - GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " - "performance"); -#else - float sumf[4][4]; - 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_0x4 * b_ptr = (const block_q4_0x4 *) 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]; - } - } - } -#endif -} - -void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * 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(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__)) - if (svcntw() == 8) { - const void * b_ptr = vx; - const void * a_ptr = vy; - float * res_ptr = s; - size_t res_stride = bs * sizeof(float); - - __asm__ __volatile__( - "mov x20, #0x4\n" - "mov x13, %x[nr]\n" - "mov z28.s, #-0x4\n" - "mov x12, #0x88\n" - "ptrue p1.b\n" - "whilelt p0.s, XZR, x20\n" - "cmp x13, #0x10\n" - "mul x12, %x[nb], x12\n" - "blt 4f\n" - "1:" // Row loop - "add x11, %x[b_ptr], #0x10\n" - "mov x10, %x[nc]\n" - "add x9, %x[res_ptr], %x[res_stride], LSL #4\n" - "2:" // Column loop - "add x28, %x[a_ptr], #0x8\n" - "mov z24.b, #0x0\n" - "mov z15.b, #0x0\n" - "mov x27, %x[nb]\n" - "add x26, x28, x12\n" - "mov z12.b, #0x0\n" - "mov z0.b, #0x0\n" - "add x25, x26, x12\n" - "mov z13.b, #0x0\n" - "mov z1.b, #0x0\n" - "add x24, x25, x12\n" - "mov z20.b, #0x0\n" - "mov z25.b, #0x0\n" - "mov z11.b, #0x0\n" - "mov z16.b, #0x0\n" - "mov z19.b, #0x0\n" - "mov z26.b, #0x0\n" - "mov z8.b, #0x0\n" - "mov z29.b, #0x0\n" - "mov z27.b, #0x0\n" - "mov z10.b, #0x0\n" - "3:" // Block loop - "ld1b { z30.b }, p1/Z, [x11]\n" - "ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n" - "mov z18.s, #0x0\n" - "mov z7.s, #0x0\n" - "ld1rqb { z3.b }, p1/Z, [x28]\n" - "ld1rqb { z5.b }, p1/Z, [x28, #16]\n" - "mov z9.s, #0x0\n" - "mov z22.s, #0x0\n" - "ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n" - "ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n" - "sub x20, x11, #0x10\n" - "sub x23, x28, #0x8\n" - "lsl z31.b, z30.b, #0x4\n" - "lsl z6.b, z21.b, #0x4\n" - "ld1h { z23.s }, p1/Z, [x20]\n" - "sub x22, x26, #0x8\n" - "and z30.b, z30.b, #0xf0\n" - "and z21.b, z21.b, #0xf0\n" - "sub x21, x25, #0x8\n" - "sub x20, x24, #0x8\n" - "lsl z14.b, z4.b, #0x4\n" - "lsl z2.b, z17.b, #0x4\n" - "subs x27, x27, #0x1\n" - "add x11, x11, #0x90\n" - ".inst 0x451f9872 // smmla z18.s, z3.b, z31.b\n" - ".inst 0x45069867 // smmla z7.s, z3.b, z6.b\n" - "ld1rqb { z3.b }, p1/Z, [x28, #32]\n" - "and z4.b, z4.b, #0xf0\n" - ".inst 0x451f98a9 // smmla z9.s, z5.b, z31.b\n" - ".inst 0x450698b6 // smmla z22.s, z5.b, z6.b\n" - "ld1rqb { z5.b }, p1/Z, [x28, #48]\n" - "and z17.b, z17.b, #0xf0\n" - "fcvt z23.s, p1/m, z23.h\n" - ".inst 0x450e9872 // smmla z18.s, z3.b, z14.b\n" - ".inst 0x45029867 // smmla z7.s, z3.b, z2.b\n" - "ld1rqb { z3.b }, p1/Z, [x28, #64]\n" - ".inst 0x450e98a9 // smmla z9.s, z5.b, z14.b\n" - ".inst 0x450298b6 // smmla z22.s, z5.b, z2.b\n" - "ld1rqb { z5.b }, p1/Z, [x28, #80]\n" - "fscale z23.s, p1/m, z23.s, z28.s\n" - ".inst 0x451e9872 // smmla z18.s, z3.b, z30.b\n" - ".inst 0x45159867 // smmla z7.s, z3.b, z21.b\n" - "ld1rqb { z3.b }, p1/Z, [x28, #96]\n" - ".inst 0x451e98a9 // smmla z9.s, z5.b, z30.b\n" - ".inst 0x451598b6 // smmla z22.s, z5.b, z21.b\n" - "ld1rqb { z5.b }, p1/Z, [x28, #112]\n" - "add x28, x28, #0x88\n" - ".inst 0x45049872 // smmla z18.s, z3.b, z4.b\n" - ".inst 0x45119867 // smmla z7.s, z3.b, z17.b\n" - "ld1h { z3.s }, p0/Z, [x23]\n" - ".inst 0x450498a9 // smmla z9.s, z5.b, z4.b\n" - ".inst 0x451198b6 // smmla z22.s, z5.b, z17.b\n" - "fcvt z3.s, p1/m, z3.h\n" - "uzp1 z5.d, z18.d, z7.d\n" - "uzp2 z18.d, z18.d, z7.d\n" - "mov z3.q, z3.q[0]\n" - "uzp1 z7.d, z9.d, z22.d\n" - "uzp2 z22.d, z9.d, z22.d\n" - "fmul z9.s, z23.s, z3.s[0]\n" - "scvtf z5.s, p1/m, z5.s\n" - "scvtf z18.s, p1/m, z18.s\n" - "scvtf z7.s, p1/m, z7.s\n" - "scvtf z22.s, p1/m, z22.s\n" - "fmla z24.s, p1/M, z5.s, z9.s\n" - "ld1rqb { z5.b }, p1/Z, [x26]\n" - "fmul z9.s, z23.s, z3.s[1]\n" - "fmla z15.s, p1/M, z18.s, z9.s\n" - "ld1rqb { z18.b }, p1/Z, [x26, #16]\n" - "fmul z9.s, z23.s, z3.s[2]\n" - "fmul z3.s, z23.s, z3.s[3]\n" - "fmla z12.s, p1/M, z7.s, z9.s\n" - "mov z9.s, #0x0\n" - "ld1h { z7.s }, p0/Z, [x22]\n" - ".inst 0x451f98a9 // smmla z9.s, z5.b, z31.b\n" - "fmla z0.s, p1/M, z22.s, z3.s\n" - "mov z22.s, #0x0\n" - "ld1h { z3.s }, p0/Z, [x21]\n" - ".inst 0x450698b6 // smmla z22.s, z5.b, z6.b\n" - "ld1rqb { z5.b }, p1/Z, [x26, #32]\n" - "fcvt z7.s, p1/m, z7.h\n" - "fcvt z3.s, p1/m, z3.h\n" - ".inst 0x450e98a9 // smmla z9.s, z5.b, z14.b\n" - ".inst 0x450298b6 // smmla z22.s, z5.b, z2.b\n" - "ld1rqb { z5.b }, p1/Z, [x26, #64]\n" - "mov z7.q, z7.q[0]\n" - "mov z3.q, z3.q[0]\n" - ".inst 0x451e98a9 // smmla z9.s, z5.b, z30.b\n" - ".inst 0x451598b6 // smmla z22.s, z5.b, z21.b\n" - "ld1rqb { z5.b }, p1/Z, [x26, #96]\n" - ".inst 0x450498a9 // smmla z9.s, z5.b, z4.b\n" - ".inst 0x451198b6 // smmla z22.s, z5.b, z17.b\n" - "uzp1 z5.d, z9.d, z22.d\n" - "scvtf z5.s, p1/m, z5.s\n" - "uzp2 z22.d, z9.d, z22.d\n" - "fmul z9.s, z23.s, z7.s[0]\n" - "scvtf z22.s, p1/m, z22.s\n" - "fmla z13.s, p1/M, z5.s, z9.s\n" - "ld1rqb { z9.b }, p1/Z, [x25]\n" - "fmul z5.s, z23.s, z7.s[1]\n" - "fmla z1.s, p1/M, z22.s, z5.s\n" - "mov z5.s, #0x0\n" - "mov z22.s, #0x0\n" - ".inst 0x451f9a45 // smmla z5.s, z18.b, z31.b\n" - ".inst 0x45069a56 // smmla z22.s, z18.b, z6.b\n" - "ld1rqb { z18.b }, p1/Z, [x26, #48]\n" - ".inst 0x450e9a45 // smmla z5.s, z18.b, z14.b\n" - ".inst 0x45029a56 // smmla z22.s, z18.b, z2.b\n" - "ld1rqb { z18.b }, p1/Z, [x26, #80]\n" - ".inst 0x451e9a45 // smmla z5.s, z18.b, z30.b\n" - ".inst 0x45159a56 // smmla z22.s, z18.b, z21.b\n" - "ld1rqb { z18.b }, p1/Z, [x26, #112]\n" - "add x26, x26, #0x88\n" - ".inst 0x45049a45 // smmla z5.s, z18.b, z4.b\n" - ".inst 0x45119a56 // smmla z22.s, z18.b, z17.b\n" - "uzp1 z18.d, z5.d, z22.d\n" - "scvtf z18.s, p1/m, z18.s\n" - "uzp2 z22.d, z5.d, z22.d\n" - "fmul z5.s, z23.s, z7.s[2]\n" - "fmul z7.s, z23.s, z7.s[3]\n" - "scvtf z22.s, p1/m, z22.s\n" - "fmla z20.s, p1/M, z18.s, z5.s\n" - "ld1rqb { z18.b }, p1/Z, [x25, #16]\n" - "ld1h { z5.s }, p0/Z, [x20]\n" - "fcvt z5.s, p1/m, z5.h\n" - "fmla z25.s, p1/M, z22.s, z7.s\n" - "mov z22.s, #0x0\n" - "mov z7.s, #0x0\n" - ".inst 0x451f9936 // smmla z22.s, z9.b, z31.b\n" - ".inst 0x45069927 // smmla z7.s, z9.b, z6.b\n" - "ld1rqb { z9.b }, p1/Z, [x25, #32]\n" - "mov z5.q, z5.q[0]\n" - ".inst 0x450e9936 // smmla z22.s, z9.b, z14.b\n" - ".inst 0x45029927 // smmla z7.s, z9.b, z2.b\n" - "ld1rqb { z9.b }, p1/Z, [x25, #64]\n" - ".inst 0x451e9936 // smmla z22.s, z9.b, z30.b\n" - ".inst 0x45159927 // smmla z7.s, z9.b, z21.b\n" - "ld1rqb { z9.b }, p1/Z, [x25, #96]\n" - ".inst 0x45049936 // smmla z22.s, z9.b, z4.b\n" - ".inst 0x45119927 // smmla z7.s, z9.b, z17.b\n" - "uzp1 z9.d, z22.d, z7.d\n" - "scvtf z9.s, p1/m, z9.s\n" - "uzp2 z22.d, z22.d, z7.d\n" - "fmul z7.s, z23.s, z3.s[0]\n" - "scvtf z22.s, p1/m, z22.s\n" - "fmla z11.s, p1/M, z9.s, z7.s\n" - "ld1rqb { z9.b }, p1/Z, [x24]\n" - "fmul z7.s, z23.s, z3.s[1]\n" - "fmla z16.s, p1/M, z22.s, z7.s\n" - "mov z22.s, #0x0\n" - "mov z7.s, #0x0\n" - ".inst 0x451f9a56 // smmla z22.s, z18.b, z31.b\n" - ".inst 0x45069a47 // smmla z7.s, z18.b, z6.b\n" - "ld1rqb { z18.b }, p1/Z, [x25, #48]\n" - ".inst 0x450e9a56 // smmla z22.s, z18.b, z14.b\n" - ".inst 0x45029a47 // smmla z7.s, z18.b, z2.b\n" - "ld1rqb { z18.b }, p1/Z, [x25, #80]\n" - ".inst 0x451e9a56 // smmla z22.s, z18.b, z30.b\n" - ".inst 0x45159a47 // smmla z7.s, z18.b, z21.b\n" - "ld1rqb { z18.b }, p1/Z, [x25, #112]\n" - "add x25, x25, #0x88\n" - ".inst 0x45049a56 // smmla z22.s, z18.b, z4.b\n" - ".inst 0x45119a47 // smmla z7.s, z18.b, z17.b\n" - "uzp1 z18.d, z22.d, z7.d\n" - "scvtf z18.s, p1/m, z18.s\n" - "uzp2 z7.d, z22.d, z7.d\n" - "fmul z22.s, z23.s, z3.s[2]\n" - "fmul z3.s, z23.s, z3.s[3]\n" - "scvtf z7.s, p1/m, z7.s\n" - "fmla z19.s, p1/M, z18.s, z22.s\n" - "ld1rqb { z18.b }, p1/Z, [x24, #16]\n" - "fmul z22.s, z23.s, z5.s[0]\n" - "fmla z26.s, p1/M, z7.s, z3.s\n" - "mov z3.s, #0x0\n" - "mov z7.s, #0x0\n" - ".inst 0x451f9923 // smmla z3.s, z9.b, z31.b\n" - ".inst 0x45069927 // smmla z7.s, z9.b, z6.b\n" - "ld1rqb { z9.b }, p1/Z, [x24, #32]\n" - ".inst 0x450e9923 // smmla z3.s, z9.b, z14.b\n" - ".inst 0x45029927 // smmla z7.s, z9.b, z2.b\n" - "mov z9.s, #0x0\n" - ".inst 0x451f9a49 // smmla z9.s, z18.b, z31.b\n" - "mov z31.s, #0x0\n" - ".inst 0x45069a5f // smmla z31.s, z18.b, z6.b\n" - "ld1rqb { z6.b }, p1/Z, [x24, #48]\n" - "ld1rqb { z18.b }, p1/Z, [x24, #64]\n" - ".inst 0x450e98c9 // smmla z9.s, z6.b, z14.b\n" - "fmul z14.s, z23.s, z5.s[1]\n" - ".inst 0x450298df // smmla z31.s, z6.b, z2.b\n" - "ld1rqb { z6.b }, p1/Z, [x24, #80]\n" - "fmul z2.s, z23.s, z5.s[2]\n" - "fmul z23.s, z23.s, z5.s[3]\n" - ".inst 0x451e9a43 // smmla z3.s, z18.b, z30.b\n" - ".inst 0x45159a47 // smmla z7.s, z18.b, z21.b\n" - "ld1rqb { z5.b }, p1/Z, [x24, #96]\n" - ".inst 0x451e98c9 // smmla z9.s, z6.b, z30.b\n" - ".inst 0x451598df // smmla z31.s, z6.b, z21.b\n" - "ld1rqb { z18.b }, p1/Z, [x24, #112]\n" - "add x24, x24, #0x88\n" - ".inst 0x450498a3 // smmla z3.s, z5.b, z4.b\n" - ".inst 0x451198a7 // smmla z7.s, z5.b, z17.b\n" - ".inst 0x45049a49 // smmla z9.s, z18.b, z4.b\n" - ".inst 0x45119a5f // smmla z31.s, z18.b, z17.b\n" - "uzp1 z18.d, z3.d, z7.d\n" - "uzp2 z5.d, z3.d, z7.d\n" - "scvtf z18.s, p1/m, z18.s\n" - "uzp1 z6.d, z9.d, z31.d\n" - "uzp2 z9.d, z9.d, z31.d\n" - "scvtf z5.s, p1/m, z5.s\n" - "fmla z8.s, p1/M, z18.s, z22.s\n" - "scvtf z6.s, p1/m, z6.s\n" - "scvtf z9.s, p1/m, z9.s\n" - "fmla z29.s, p1/M, z5.s, z14.s\n" - "fmla z27.s, p1/M, z6.s, z2.s\n" - "fmla z10.s, p1/M, z9.s, z23.s\n" - "bgt 3b\n" - "mov x20, %x[res_ptr]\n" - "subs x10, x10, #0x8\n" - "add %x[res_ptr], %x[res_ptr], #0x20\n" - "st1w { z24.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z15.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z12.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z0.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z13.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z1.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z20.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z25.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z11.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z16.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z19.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z26.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z8.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z29.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z27.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "st1w { z10.s }, p1, [x20]\n" - "bne 2b\n" - "mov x20, #0x4\n" - "sub x13, x13, #0x10\n" - "cmp x13, #0x10\n" - "mov %x[res_ptr], x9\n" - "madd %x[a_ptr], x20, x12, %x[a_ptr]\n" - "bge 1b\n" - "4:" // Row loop skip - "cbz x13, 9f\n" - "5:" // Row tail: Row loop - "add x25, %x[b_ptr], #0x10\n" - "mov x24, %x[nc]\n" - "add x23, %x[res_ptr], %x[res_stride], LSL #2\n" - "6:" // Row tail: Column loop - "mov z24.b, #0x0\n" - "mov z15.b, #0x0\n" - "add x28, %x[a_ptr], #0x8\n" - "mov x22, %x[nb]\n" - "mov z12.b, #0x0\n" - "mov z0.b, #0x0\n" - "7:" // Row tail: Block loop - "ld1b { z3.b }, p1/Z, [x25]\n" - "ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n" - "mov z2.s, #0x0\n" - "mov z25.s, #0x0\n" - "ld1rqb { z26.b }, p1/Z, [x28]\n" - "ld1rqb { z21.b }, p1/Z, [x28, #16]\n" - "mov z27.s, #0x0\n" - "mov z19.s, #0x0\n" - "ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n" - "ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n" - "sub x21, x25, #0x10\n" - "sub x20, x28, #0x8\n" - "lsl z20.b, z3.b, #0x4\n" - "lsl z4.b, z6.b, #0x4\n" - "ld1rqb { z10.b }, p1/Z, [x28, #32]\n" - "ld1rqb { z23.b }, p1/Z, [x28, #48]\n" - "and z3.b, z3.b, #0xf0\n" - "and z6.b, z6.b, #0xf0\n" - "ld1rqb { z11.b }, p1/Z, [x28, #64]\n" - "ld1rqb { z7.b }, p1/Z, [x28, #80]\n" - "lsl z8.b, z29.b, #0x4\n" - "lsl z14.b, z16.b, #0x4\n" - "ld1rqb { z18.b }, p1/Z, [x28, #96]\n" - "ld1rqb { z30.b }, p1/Z, [x28, #112]\n" - ".inst 0x45149b42 // smmla z2.s, z26.b, z20.b\n" - ".inst 0x45049b59 // smmla z25.s, z26.b, z4.b\n" - "and z29.b, z29.b, #0xf0\n" - "ld1h { z17.s }, p1/Z, [x21]\n" - ".inst 0x45149abb // smmla z27.s, z21.b, z20.b\n" - ".inst 0x45049ab3 // smmla z19.s, z21.b, z4.b\n" - "and z16.b, z16.b, #0xf0\n" - "ld1h { z4.s }, p0/Z, [x20]\n" - "subs x22, x22, #0x1\n" - "add x28, x28, #0x88\n" - "fcvt z17.s, p1/m, z17.h\n" - "add x25, x25, #0x90\n" - ".inst 0x45089942 // smmla z2.s, z10.b, z8.b\n" - ".inst 0x450e9959 // smmla z25.s, z10.b, z14.b\n" - "fcvt z4.s, p1/m, z4.h\n" - ".inst 0x45089afb // smmla z27.s, z23.b, z8.b\n" - ".inst 0x450e9af3 // smmla z19.s, z23.b, z14.b\n" - "fscale z17.s, p1/m, z17.s, z28.s\n" - "mov z4.q, z4.q[0]\n" - ".inst 0x45039962 // smmla z2.s, z11.b, z3.b\n" - ".inst 0x45069979 // smmla z25.s, z11.b, z6.b\n" - "fmul z23.s, z17.s, z4.s[0]\n" - "fmul z9.s, z17.s, z4.s[1]\n" - "fmul z21.s, z17.s, z4.s[2]\n" - "fmul z4.s, z17.s, z4.s[3]\n" - ".inst 0x450398fb // smmla z27.s, z7.b, z3.b\n" - ".inst 0x450698f3 // smmla z19.s, z7.b, z6.b\n" - ".inst 0x451d9a42 // smmla z2.s, z18.b, z29.b\n" - ".inst 0x45109a59 // smmla z25.s, z18.b, z16.b\n" - ".inst 0x451d9bdb // smmla z27.s, z30.b, z29.b\n" - ".inst 0x45109bd3 // smmla z19.s, z30.b, z16.b\n" - "uzp1 z31.d, z2.d, z25.d\n" - "uzp2 z13.d, z2.d, z25.d\n" - "scvtf z31.s, p1/m, z31.s\n" - "uzp1 z17.d, z27.d, z19.d\n" - "uzp2 z18.d, z27.d, z19.d\n" - "scvtf z13.s, p1/m, z13.s\n" - "fmla z24.s, p1/M, z31.s, z23.s\n" - "scvtf z17.s, p1/m, z17.s\n" - "scvtf z18.s, p1/m, z18.s\n" - "fmla z15.s, p1/M, z13.s, z9.s\n" - "fmla z12.s, p1/M, z17.s, z21.s\n" - "fmla z0.s, p1/M, z18.s, z4.s\n" - "bgt 7b\n" - "mov x20, %x[res_ptr]\n" - "cmp x13, #0x1\n" - "st1w { z24.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x13, #0x2\n" - "st1w { z15.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "cmp x13, #0x3\n" - "st1w { z12.s }, p1, [x20]\n" - "add x20, x20, %x[res_stride]\n" - "ble 8f\n" - "st1w { z0.s }, p1, [x20]\n" - "8:" // Row tail: Accumulator store skip - "subs x24, x24, #0x8\n" - "add %x[res_ptr], %x[res_ptr], #0x20\n" - "bne 6b\n" - "subs x13, x13, #0x4\n" - "add %x[a_ptr], %x[a_ptr], x12\n" - "mov %x[res_ptr], x23\n" - "bgt 5b\n" - "9:" // Row tail: Row loop skip - : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr) - : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc) - : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" - ); - return; - } - else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { - GGML_ASSERT((ggml_cpu_has_sve() && (svcntw() == 8)) && - "__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal " - "performance"); - } - else if (ggml_cpu_has_neon()) { - GGML_ASSERT(((ggml_cpu_has_sve() && (svcntw() == 8)) || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 " - "quantization format for optimal performance"); - } -#endif -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) - GGML_ASSERT(ggml_cpu_has_sve() && - "__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance"); -#elif defined(__ARM_NEON) && defined(__aarch64__) - GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) && - "__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal " - "performance"); -#else - 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]; - } - } - } -#endif -} diff --git a/ggml/src/ggml-aarch64.h b/ggml/src/ggml-aarch64.h deleted file mode 100644 index 517babaf..00000000 --- a/ggml/src/ggml-aarch64.h +++ /dev/null @@ -1,39 +0,0 @@ -// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd. -#pragma once - -#define GGML_COMMON_DECL_C -#include "ggml-common.h" - -#include "ggml.h" - -// GGML internal header - -#ifdef __cplusplus -extern "C" { -#endif - -// Quantization -void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); -void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k); - -void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave); - -// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization") -size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); -size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); -size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix); - -// GEMV -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); - -// GEMM -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); - -#ifdef __cplusplus -} -#endif - diff --git a/ggml/src/ggml-amx.cpp b/ggml/src/ggml-amx.cpp deleted file mode 100644 index ac6ec234..00000000 --- a/ggml/src/ggml-amx.cpp +++ /dev/null @@ -1,453 +0,0 @@ -#include "ggml-amx.h" -#include "ggml-amx/common.h" -#include "ggml-amx/mmq.h" -#include "ggml-backend-impl.h" -#include "ggml-impl.h" - -#if defined(__gnu_linux__) -#include -#include -#endif - -#include -#include -#include - -#if defined(__AMX_INT8__) - -// AMX buffer interface -static const char * ggml_backend_amx_buffer_get_name(ggml_backend_buffer_t buffer) { - return "AMX"; - - GGML_UNUSED(buffer); -} - -static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) { - free(buffer->context); -} - -static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) { - return (void *)(buffer->context); -} - -static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) { - memset((char *)tensor->data + offset, value, size); - - GGML_UNUSED(buffer); -} - -static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - if (qtype_has_amx_kernels(tensor->type)) { - ggml_backend_amx_convert_weight(tensor, data, offset, size); - } else { - memcpy((char *)tensor->data + offset, data, size); - } - - GGML_UNUSED(buffer); -} - -static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { - GGML_ASSERT(!qtype_has_amx_kernels(tensor->type)); - memcpy(data, (const char *)tensor->data + offset, size); - - GGML_UNUSED(buffer); -} - -static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) { - if (ggml_backend_buffer_is_host(src->buffer)) { - if (qtype_has_amx_kernels(src->type)) { - ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_backend_amx_get_alloc_size(dst)); - } else { - memcpy(dst->data, src->data, ggml_nbytes(src)); - } - return true; - } - return false; - - GGML_UNUSED(buffer); -} - -static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) { - memset(buffer->context, value, buffer->size); -} - -static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = { - /* .get_name = */ ggml_backend_amx_buffer_get_name, - /* .free_buffer = */ ggml_backend_amx_buffer_free_buffer, - /* .get_base = */ ggml_backend_amx_buffer_get_base, - /* .init_tensor = */ NULL, // no initialization required - /* .memset_tensor = */ ggml_backend_amx_buffer_memset_tensor, - /* .set_tensor = */ ggml_backend_amx_buffer_set_tensor, - /* .get_tensor = */ ggml_backend_amx_buffer_get_tensor, - /* .cpy_tensor = */ ggml_backend_amx_buffer_cpy_tensor, - /* .clear = */ ggml_backend_amx_buffer_clear, - /* .reset = */ NULL, -}; - -static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) { - return "AMX"; - - GGML_UNUSED(buft); -} - -static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { - void * data = aligned_alloc(TENSOR_ALIGNMENT, size); - if (data == NULL) { - fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size); - return NULL; - } - - return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size); -} - -static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { - return TENSOR_ALIGNMENT; - - GGML_UNUSED(buft); -} - -static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) { - return ggml_backend_amx_get_alloc_size(tensor); - - GGML_UNUSED(buft); -} - -static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) { - return false; - - GGML_UNUSED(buft); -} - -ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() { - static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = { - /* .iface = */ { - /* .get_name = */ ggml_backend_amx_buffer_type_get_name, - /* .alloc_buffer = */ ggml_backend_amx_buffer_type_alloc_buffer, - /* .get_alignment = */ ggml_backend_amx_buffer_type_get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX - /* .get_alloc_size = */ ggml_backend_amx_buffer_type_get_alloc_size, - /* .is_host = */ ggml_backend_amx_buffer_type_is_host, - }, - /* .device = */ NULL, - /* .context = */ NULL, - }; - - return &ggml_backend_buffer_type_amx; -} - -// backend interface - -static const char * ggml_backend_amx_name(ggml_backend_t backend) { - return "AMX"; - - GGML_UNUSED(backend); -} - -static void ggml_backend_amx_free(ggml_backend_t backend) { - ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context; - delete ctx; - delete backend; -} - -static ggml_backend_buffer_type_t ggml_backend_amx_get_default_buffer_type(ggml_backend_t backend) { - return ggml_backend_amx_buffer_type(); - - GGML_UNUSED(backend); -} - -static enum ggml_status ggml_backend_amx_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) { - ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend->context; - - for (int i = 0; i < cgraph->n_nodes; i++) { - struct ggml_tensor * node = cgraph->nodes[i]; - - switch (node->op) { - case GGML_OP_MUL_MAT: - ggml_backend_amx_mul_mat(ctx, node); - break; - - case GGML_OP_NONE: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_PERMUTE: - case GGML_OP_TRANSPOSE: - break; - - default: - fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node)); - GGML_ASSERT(false); - } - } - - return GGML_STATUS_SUCCESS; - - GGML_UNUSED(backend); -} - -static struct ggml_backend_i ggml_backend_amx_i = { - /* .get_name = */ ggml_backend_amx_name, - /* .free = */ ggml_backend_amx_free, - /* .get_default_buffer_type = */ ggml_backend_amx_get_default_buffer_type, - /* .set_tensor_async = */ NULL, - /* .get_tensor_async = */ NULL, - /* .cpy_tensor_async = */ NULL, - /* .synchronize = */ NULL, - /* .graph_plan_create = */ NULL, - /* .graph_plan_free = */ NULL, - /* .graph_plan_update = */ NULL, - /* .graph_plan_compute = */ NULL, - /* .graph_compute = */ ggml_backend_amx_graph_compute, - /* .supports_op = */ NULL, - /* .supports_buft = */ NULL, - /* .offload_op = */ NULL, - /* .event_record = */ NULL, - /* .event_wait = */ NULL, -}; - -static ggml_guid_t ggml_backend_amx_guid() { - static ggml_guid guid = { 0x13, 0xb8, 0xa4, 0xc4, 0xba, 0xfe, 0x51, 0x67, 0x87, 0x44, 0x55, 0x15, 0xb2, 0x35, 0x62, 0x3e }; - return &guid; -} - -#define ARCH_GET_XCOMP_PERM 0x1022 -#define ARCH_REQ_XCOMP_PERM 0x1023 -#define XFEATURE_XTILECFG 17 -#define XFEATURE_XTILEDATA 18 - -static bool ggml_amx_init() { -#if defined(__gnu_linux__) - if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) { - fprintf(stderr, "AMX is not ready to be used!\n"); - return false; - } - return true; -#elif defined(_WIN32) - return true; -#endif -} - -ggml_backend_t ggml_backend_amx_init() { - - // invoke a Linux system call to request access to AMX features - ggml_amx_init(); - - // backend context - ggml_backend_amx_context * ctx = new ggml_backend_amx_context; - - // ggml amx backend - ggml_backend_t backend = new ggml_backend { - /* .guid = */ ggml_backend_amx_guid(), - /* .interface = */ ggml_backend_amx_i, - /* .device = */ ggml_backend_reg_dev_get(ggml_backend_amx_reg(), 0), - /* .context = */ ctx, - }; - - return backend; -} - -bool ggml_backend_is_amx(ggml_backend_t backend) { - return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_amx_guid()); -} - -void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) { - GGML_ASSERT(ggml_backend_is_amx(backend_amx)); - - ggml_backend_amx_context * ctx = (ggml_backend_amx_context *)backend_amx->context; - ctx->n_threads = n_threads; -} - -// device interface - -static const char * ggml_backend_amx_device_get_name(ggml_backend_dev_t dev) { - return "AMX"; - - GGML_UNUSED(dev); -} - -static const char * ggml_backend_amx_device_get_description(ggml_backend_dev_t dev) { - return "Intel Advanced Matrix Extensions"; - - GGML_UNUSED(dev); -} - -static void ggml_backend_amx_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) { - // TODO - *free = 0; - *total = 0; - - GGML_UNUSED(dev); -} - -static enum ggml_backend_dev_type ggml_backend_amx_device_get_type(ggml_backend_dev_t dev) { - return GGML_BACKEND_DEVICE_TYPE_CPU; - - GGML_UNUSED(dev); -} - -static void ggml_backend_amx_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) { - props->name = ggml_backend_amx_device_get_name(dev); - props->description = ggml_backend_amx_device_get_description(dev); - props->type = ggml_backend_amx_device_get_type(dev); - ggml_backend_amx_device_get_memory(dev, &props->memory_free, &props->memory_total); - - // `buffer_from_host_ptr` is intended to be used in mmap, when memory layout unchanged - props->caps = { - /* .async = */ false, - /* .host_buffer = */ false, - /* .buffer_from_host_ptr = */ false, - /* .events = */ false, - }; -} - -static ggml_backend_t ggml_backend_amx_device_init(ggml_backend_dev_t dev, const char * params) { - return ggml_backend_amx_init(); - - GGML_UNUSED(dev); - GGML_UNUSED(params); -} - -static ggml_backend_buffer_type_t ggml_backend_amx_device_get_buffer_type(ggml_backend_dev_t dev) { - return ggml_backend_amx_buffer_type(); - - GGML_UNUSED(dev); -} - -static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) { - - // handle only 2d gemm for now - auto is_contiguous_2d = [](const struct ggml_tensor * t) { - return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1; - }; - - switch (op->op) { - case GGML_OP_NONE: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_PERMUTE: - case GGML_OP_TRANSPOSE: - return true; - - case GGML_OP_MUL_MAT: { - const struct ggml_tensor * src0 = op->src[0]; - const struct ggml_tensor * src1 = op->src[1]; - - const enum ggml_type type = src0->type; - const int64_t ne0 = op->ne[0]; - - bool is_training = src0->grad || src1->grad; - - // amx kernels enables for Q4_0, Q4_1, Q8_0, F16 - // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256 - bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16); - - bool can_use_amx = - is_contiguous_2d(src0) && // src0 must be contiguous - is_contiguous_2d(src1) && // src1 must be contiguous - !is_training && // inference only - src1->type == GGML_TYPE_F32 && // src1 must be float32 - has_amx_kernels && // with amx kernel impls - ne0 % (TILE_N * 2) == 0; // out_features is 32x - - return can_use_amx; - } - default: - return false; - } - - GGML_UNUSED(dev); -} - -static bool ggml_backend_amx_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) { - return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name; - - GGML_UNUSED(dev); -} - -static const struct ggml_backend_device_i ggml_backend_amx_device_i = { - /* .get_name = */ ggml_backend_amx_device_get_name, - /* .get_description = */ ggml_backend_amx_device_get_description, - /* .get_memory = */ ggml_backend_amx_device_get_memory, - /* .get_type = */ ggml_backend_amx_device_get_type, - /* .get_props = */ ggml_backend_amx_device_get_props, - /* .init_backend = */ ggml_backend_amx_device_init, - /* .get_buffer_type = */ ggml_backend_amx_device_get_buffer_type, - /* .get_host_buffer_type = */ NULL, - /* .buffer_from_host_ptr = */ NULL, - /* .supports_op = */ ggml_backend_amx_device_supports_op, - /* .supports_buft = */ ggml_backend_amx_device_supports_buft, - /* .offload_op = */ NULL, - /* .event_new = */ NULL, - /* .event_free = */ NULL, - /* .event_synchronize = */ NULL, -}; - -// backend reg interface - -static const char * ggml_backend_amx_reg_get_name(ggml_backend_reg_t reg) { - return "AMX"; - - GGML_UNUSED(reg); -} - -static size_t ggml_backend_amx_reg_get_device_count(ggml_backend_reg_t reg) { - return 1; - - GGML_UNUSED(reg); -} - -static ggml_backend_dev_t ggml_backend_amx_reg_get_device(ggml_backend_reg_t reg, size_t index) { - GGML_ASSERT(index == 0); - - static ggml_backend_device ggml_backend_amx_device = { - /* .iface = */ ggml_backend_amx_device_i, - /* .reg = */ reg, - /* .context = */ nullptr, - }; - - return &ggml_backend_amx_device; - - GGML_UNUSED(reg); - GGML_UNUSED(index); -} - -static void * ggml_backend_amx_get_proc_address(ggml_backend_reg_t reg, const char * name) { - if (std::strcmp(name, "ggml_backend_set_n_threads") == 0) { - return (void *)ggml_backend_amx_set_n_threads; - } - return NULL; - - GGML_UNUSED(reg); - GGML_UNUSED(name); -} - -static const struct ggml_backend_reg_i ggml_backend_amx_reg_i = { - /* .get_name = */ ggml_backend_amx_reg_get_name, - /* .get_device_count = */ ggml_backend_amx_reg_get_device_count, - /* .get_device = */ ggml_backend_amx_reg_get_device, - /* .get_proc_address = */ ggml_backend_amx_get_proc_address, -}; - -ggml_backend_reg_t ggml_backend_amx_reg(void) { - static struct ggml_backend_reg ggml_backend_amx_reg = { - /* .iface = */ ggml_backend_amx_reg_i, - /* .context = */ NULL, - }; - - return &ggml_backend_amx_reg; -} - -#else // if defined(__AMX_INT8__) - -ggml_backend_t ggml_backend_amx_init(void) { - fprintf(stderr, "GGML is not compiled with AMX support!\n"); - return ggml_backend_t{}; -} - -void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) { - fprintf(stderr, "GGML is not compiled with AMX support!\n"); - - GGML_UNUSED(backend_amx); - GGML_UNUSED(n_threads); -} - -#endif diff --git a/ggml/src/ggml-cann.cpp b/ggml/src/ggml-cann.cpp deleted file mode 100644 index 06930ba2..00000000 --- a/ggml/src/ggml-cann.cpp +++ /dev/null @@ -1,2020 +0,0 @@ -/* - * Copyright (c) 2023-2024 The ggml authors - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in - * all copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING - * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS - * IN THE SOFTWARE. - */ - -#include "ggml-cann.h" - -#include -#include - -#include -#include -#include -#include - -#include "ggml-backend-impl.h" -#include "ggml-cann/aclnn_ops.h" -#include "ggml-cann/common.h" - -#define GGML_COMMON_DECL_C - -#include "ggml-common.h" - -/** - * @brief Default logging callback for GGML. - * - * This function is the default logging callback that logs messages to stderr. - * - * @param level The log level. - * @param msg The log message. - * @param user_data User data passed to the callback. - */ -static void ggml_cann_default_log_callback(enum ggml_log_level level, - const char* msg, void* user_data) { - GGML_UNUSED(level); - GGML_UNUSED(user_data); - fprintf(stderr, "%s", msg); -} - -ggml_log_callback ggml_cann_log_callback = ggml_cann_default_log_callback; -void* ggml_cann_log_user_data = NULL; - -GGML_API void ggml_backend_cann_log_set_callback(ggml_log_callback log_callback, - void* user_data) { - ggml_cann_log_callback = log_callback; - ggml_cann_log_user_data = user_data; -} - -#define GGML_CANN_LOG_INFO(...) ggml_cann_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__) -#define GGML_CANN_LOG_WARN(...) ggml_cann_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__) -#define GGML_CANN_LOG_ERROR(...) \ - ggml_cann_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__) - -GGML_ATTRIBUTE_FORMAT(2, 3) - -/** - * @brief Log a message using the current logging callback. - * - * This function formats a log message and passes it to the current logging - * callback. - * - * @param level The log level. - * @param format The format string for the log message. - * @param ... The arguments for the format string. - */ -static void ggml_cann_log(enum ggml_log_level level, const char* format, ...) { - if (ggml_cann_log_callback != NULL) { - va_list args; - va_start(args, format); - char buffer[128]; - int len = vsnprintf(buffer, 128, format, args); - if (len < 128) { - ggml_cann_log_callback(level, buffer, ggml_cann_log_user_data); - } else { - // vsnprintf adds a null terminator - std::vector buffer2(len + 1); - va_end(args); - va_start(args, format); - vsnprintf(&buffer2[0], buffer2.size(), format, args); - ggml_cann_log_callback(level, buffer2.data(), - ggml_cann_log_user_data); - } - va_end(args); - } -} - -/** - * @brief Handles CANN errors by printing an error message and aborting. - * - * @param stmt The statement that caused the error. - * @param func The function in which the error occurred. - * @param file The file in which the error occurred. - * @param line The line number where the error occurred. - * @param msg The error message. - */ -[[noreturn]] void ggml_cann_error(const char* stmt, const char* func, - const char* file, int line, const char* msg) { - int32_t id = -1; - aclrtGetDevice(&id); - - GGML_CANN_LOG_ERROR("CANN error: %s\n", msg); - GGML_CANN_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func, - file, line); - GGML_CANN_LOG_ERROR(" %s\n", stmt); - // abort with GGML_ASSERT to get a stack trace - GGML_ABORT("CANN error"); -} - -/** - * @brief Sets the device to be used by CANN. - * - * @param device The device ID to set. - */ -void ggml_cann_set_device(const int32_t device) { - // TODO: uncomment these lines after empty context has fixed. - // int current_device; - // ACL_CHECK(aclrtGetDevice(¤t_device)); - - // if (device == current_device) { - // return; - // } - ACL_CHECK(aclrtSetDevice(device)); -} - -/** - * @brief Retrieves the current device ID. - * - * @return The current device ID. - */ -int32_t ggml_cann_get_device() { - int32_t id; - ACL_CHECK(aclrtGetDevice(&id)); - return id; -} - -/** - * @brief Initialize the CANN device information. - * - * This function initializes the CANN device information by obtaining the - * device count and setting the memory allocation granularity for each device. - * - * @return A structure containing the device information. - */ -static ggml_cann_device_info ggml_cann_init() { - ggml_cann_device_info info = {}; - - aclError err = aclrtGetDeviceCount((uint32_t*)&info.device_count); - - if (err != ACL_SUCCESS) { - GGML_CANN_LOG_ERROR("%s: failed to initialize CANN: %s\n", - __func__, aclGetRecentErrMsg()); - return info; - } - - GGML_ASSERT(info.device_count <= GGML_CANN_MAX_DEVICES); - - for (int id = 0; id < info.device_count; ++id) { - aclrtPhysicalMemProp prop = {}; - prop.handleType = ACL_MEM_HANDLE_TYPE_NONE; - prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED; - prop.memAttr = ACL_HBM_MEM_HUGE; - prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE; - prop.location.id = id; - prop.reserve = 0; - ACL_CHECK(aclrtMemGetAllocationGranularity( - &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED, - &info.devices[id].vmm_granularity)); - } - - // TODO: add more device info later. - return info; -} - -/** - * @brief Retrieve the CANN device information. - * - * This function returns a reference to a structure containing the CANN device - * information. The device information is initialized once and reused on - * subsequent calls. - * - * @return A reference to the structure containing the device information. - */ -const ggml_cann_device_info& ggml_cann_info() { - static ggml_cann_device_info info = ggml_cann_init(); - return info; -} - -//#define DEBUG_CANN_MALLOC -/** - * @brief A pool of CANN buffers(legacy). - * - * This class manages a pool of CANN buffers for a specific device. - */ -struct ggml_cann_pool_leg : public ggml_cann_pool { - /** - * @brief The maximum number of buffers in the pool. - */ - static const int MAX_BUFFERS = 256; - - /** - * @brief The device ID associated with this buffer pool. - */ - int device; - - /** - * @brief Structure representing a CANN buffer. - */ - struct ggml_cann_buffer { - void* ptr = nullptr; ///< Pointer to the buffer memory. - size_t size = 0; ///< Size of the buffer. - }; - - /** - * @brief Array of CANN buffers in the pool. - */ - ggml_cann_buffer buffer_pool[MAX_BUFFERS] = {}; - - /** - * @brief Total size of all buffers in the pool. - */ - size_t pool_size = 0; - - /** - * @brief Constructor to initialize the buffer pool for a specific device. - * - * @param device The device ID to associate with this buffer pool. - */ - explicit ggml_cann_pool_leg(int device) : device(device) {} - - /** - * @brief Destructor to free all buffers in the pool. - */ - ~ggml_cann_pool_leg() { - ggml_cann_set_device(device); - for (int i = 0; i < MAX_BUFFERS; ++i) { - ggml_cann_buffer& b = buffer_pool[i]; - if (b.ptr != nullptr) { - ACL_CHECK(aclrtFree(b.ptr)); - pool_size -= b.size; - } - } - GGML_ASSERT(pool_size == 0); - } - - /** - * @brief Allocate a buffer of the given size. - * - * @param size The size of the buffer to allocate. - * @param actual_size A pointer to a variable to receive the actual size of - * the allocated buffer. - * @return A pointer to the allocated buffer. - */ - void* alloc(size_t size, size_t* actual_size) override { -#ifdef DEBUG_CANN_MALLOC - int nnz = 0; - size_t max_size = 0; -#endif - size_t best_diff = 1ull << 36; - int ibest = -1; - for (int i = 0; i < MAX_BUFFERS; ++i) { - ggml_cann_buffer& b = buffer_pool[i]; - if (b.ptr != nullptr) { -#ifdef DEBUG_CANN_MALLOC - ++nnz; - if (b.size > max_size) max_size = b.size; -#endif - if (b.size >= size) { - size_t diff = b.size - size; - if (diff < best_diff) { - best_diff = diff; - ibest = i; - if (!best_diff) { - void* ptr = b.ptr; - *actual_size = b.size; - b.ptr = nullptr; - b.size = 0; - return ptr; - } - } - } - } - } - if (ibest >= 0) { - ggml_cann_buffer& b = buffer_pool[ibest]; - void* ptr = b.ptr; - *actual_size = b.size; - b.ptr = nullptr; - b.size = 0; - return ptr; - } - void* ptr; - size_t look_ahead_size = (size_t)(1.05 * size); - look_ahead_size = 256 * ((look_ahead_size + 255) / 256); - ggml_cann_set_device(device); - ACL_CHECK( - aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST)); - *actual_size = look_ahead_size; - pool_size += look_ahead_size; -#ifdef DEBUG_CANN_MALLOC - GGML_CANN_LOG_INFO( - "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, " - "requested %u MB\n", - __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024), - (uint32_t)(pool_size / 1024 / 1024), - (uint32_t)(size / 1024 / 1024)); -#endif - return ptr; - } - - /** - * @brief Free a buffer and return it to the pool. - * - * @param ptr Pointer to the buffer to free. - * @param size Size of the buffer to free. - */ - void free(void* ptr, size_t size) override { - for (int i = 0; i < MAX_BUFFERS; ++i) { - ggml_cann_buffer& b = buffer_pool[i]; - if (b.ptr == nullptr) { - b.ptr = ptr; - b.size = size; - return; - } - } - // memory should always buffered. these memory may still needed by - // tasks in stream. - // TODO, fix me. - GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n"); - } -}; - -/** - * @brief A pool of CANN buffers with virtual memory. - * - * This class manages a pool of CANN buffers with virtual memory for a specific - * device. - */ -struct ggml_cann_pool_vmm : public ggml_cann_pool { - /** - * @brief The maximum size of the virtual memory pool (32 GB). - */ - static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB - - /** - * @brief The device ID associated with this buffer pool. - */ - int device; - - /** - * @brief Pointer to the start of the virtual memory pool. - */ - void* pool_addr = 0; - - /** - * @brief Amount of virtual memory used in the pool. - */ - size_t pool_used = 0; - - /** - * @brief Total size of the virtual memory pool. - */ - size_t pool_size = 0; - - /** - * @brief Allocation granularity for the virtual memory pool. - */ - size_t granularity; - - /** - * @brief Handles for the physical memory allocated. - */ - std::vector handles; - - /** - * @brief Offsets for the mapped memory regions. - */ - std::vector map_offsets; - - /** - * @brief Constructor to initialize the buffer pool with virtual memory for - * a specific device. - * - * @param device The device ID to associate with this buffer pool. - */ - explicit ggml_cann_pool_vmm(int device) - : device(device), - granularity(ggml_cann_info().devices[device].vmm_granularity) {} - - /** - * @brief Destructor to free all buffers in the virtual memory pool. - */ - ~ggml_cann_pool_vmm() { - if (pool_addr != 0) { - for (auto& offset : map_offsets) { - ACL_CHECK(aclrtUnmapMem(offset)); - } - for (auto& handle : handles) { - ACL_CHECK(aclrtFreePhysical(handle)); - } - ACL_CHECK(aclrtReleaseMemAddress(pool_addr)); - } - } - - /** - * @brief Allocate a buffer of the given size in the virtual memory pool. - * - * @param size The size of the buffer to allocate. - * @param actual_size A pointer to a variable to receive the actual size of - * the allocated buffer. - * @return A pointer to the allocated buffer. - */ - void* alloc(size_t size, size_t* actual_size) override { - // round up the allocation size to the alignment to ensure that all - // allocations are aligned for all data types - const size_t alignment = 128; - size = alignment * ((size + alignment - 1) / alignment); - - size_t avail = pool_size - pool_used; - - if (size > avail) { - // round up to the next multiple of the granularity - size_t reserve_size = size - avail; - reserve_size = - granularity * ((reserve_size + granularity - 1) / granularity); - - GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE); - - // allocate more physical memory - aclrtPhysicalMemProp prop = {}; - prop.handleType = ACL_MEM_HANDLE_TYPE_NONE; - prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED; - prop.memAttr = ACL_HBM_MEM_HUGE; - prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE; - prop.location.id = device; - prop.reserve = 0; - aclrtDrvMemHandle handle; - ACL_CHECK(aclrtMallocPhysical(&handle, reserve_size, &prop, 0)); - - // reserve virtual address space (if not already reserved) - if (pool_addr == 0) { - ACL_CHECK(aclrtReserveMemAddress( - &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1)); - } - - // map at the end of the pool - ACL_CHECK(aclrtMapMem((char*)pool_addr + pool_size, reserve_size, 0, - handle, 0)); - - handles.push_back(handle); - map_offsets.push_back((char*)pool_addr + pool_size); - - // add to the pool - pool_size += reserve_size; - - // GGML_CANN_LOG_INFO("cann pool[%d]: size increased to %llu MB ( - // reserved %llu MB)\n", - // device, (unsigned long long) (pool_size/1024/1024), - // (unsigned long long) (reserve_size/1024/1024)); - } - - GGML_ASSERT(pool_addr != 0); - - void* ptr = (void*)((char*)pool_addr + pool_used); - *actual_size = size; - pool_used += size; - -#ifdef DEBUG_CANN_MALLOC - GGML_CANN_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device, - (unsigned long long)size, (unsigned long long)ptr); -#endif - return ptr; - } - - /** - * @brief Free a buffer and return it to the virtual memory pool. - * - * @param ptr Pointer to the buffer to free. - * @param size Size of the buffer to free. - */ - void free(void* ptr, size_t size) override { -#ifdef DEBUG_CANN_MALLOC - GGML_CANN_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device, - (unsigned long long)size, (unsigned long long)ptr); -#endif - - pool_used -= size; - - // all deallocations must be in reverse order of the allocations - GGML_ASSERT(ptr == (void*)((char*)pool_addr + pool_used)); - } -}; - -/** - * @brief Create a new CANN pool for a specific device. - * - * Factory method to create a new CANN pool object based on the device type. - * - * @param device The device ID for which to create the pool. - * @return A unique pointer to the created CANN pool. - */ -std::unique_ptr ggml_backend_cann_context::new_pool_for_device( - int device) { - // return std::unique_ptr(new ggml_cann_pool_leg(device)); - return std::unique_ptr(new ggml_cann_pool_vmm(device)); -} - -// cann buffer -/** - * @brief Context for managing a CANN buffer associated with a specific device. - * - * This structure holds information about a CANN buffer, including the device - * ID, device pointer, and a name derived from GGML_CANN_NAME and the device ID. - */ -struct ggml_backend_cann_buffer_context { - int32_t device; ///< The device ID associated with this buffer context. - void* dev_ptr = - nullptr; ///< Pointer to the device memory allocated for the buffer. - - /** - * @brief Constructor to initialize the CANN buffer context. - * - * @param device The device ID associated with this buffer context. - * @param dev_ptr Pointer to the device memory allocated for the buffer. - */ - ggml_backend_cann_buffer_context(int32_t device, void* dev_ptr) - : device(device), - dev_ptr(dev_ptr) {} - - /** - * @brief Destructor to free the device memory allocated for the buffer. - */ - ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); } -}; - -/** - * @brief Retrieve the name associated with a CANN buffer. - * - * This function returns the name of a CANN buffer, which is stored in the - * context of the buffer. - * - * @param buffer The CANN buffer whose name is to be retrieved. - * @return A pointer to a C-string containing the name of the buffer. - */ - -GGML_CALL static const char* ggml_backend_cann_buffer_get_name( - ggml_backend_buffer_t buffer) { - return "CANN"; - - GGML_UNUSED(buffer); -} - -/** - * @brief Check if a buffer is a CANN buffer. - * - * This function checks if a given buffer is a CANN buffer by comparing its - * `get_name` function pointer to `ggml_backend_cann_buffer_get_name`. - * - * @param buffer The buffer to check. - * @return true if the buffer is a CANN buffer, false otherwise. - */ -GGML_CALL static bool ggml_backend_buffer_is_cann( - ggml_backend_buffer_t buffer) { - return buffer->iface.get_name == ggml_backend_cann_buffer_get_name; -} - -/** - * @brief Free resources associated with a CANN buffer. - * - * This function frees the resources associated with a CANN buffer, including - * its context. - * - * @param buffer The CANN buffer to free. - */ -GGML_CALL static void ggml_backend_cann_buffer_free_buffer( - ggml_backend_buffer_t buffer) { - ggml_backend_cann_buffer_context* ctx = - (ggml_backend_cann_buffer_context*)buffer->context; - delete ctx; -} - -/** - * @brief Retrieve the base pointer of a CANN buffer. - * - * This function returns the base pointer of a CANN buffer, which points to the - * device memory allocated for the buffer. - * - * @param buffer The CANN buffer whose base pointer is to be retrieved. - * @return A pointer to the base of the device memory allocated for the buffer. - */ -GGML_CALL static void* ggml_backend_cann_buffer_get_base( - ggml_backend_buffer_t buffer) { - ggml_backend_cann_buffer_context* ctx = - (ggml_backend_cann_buffer_context*)buffer->context; - return ctx->dev_ptr; -} - -/** - * @brief Transform quantized Q4.0 tensor data into a format suitable for CANN - * processing. - * - * This function transforms quantized Q4.0 tensor data into a format suitable - * for CANN processing. It extracts quantization values and scales from the - * source data and prepares them in a format expected by CANN operations. - * - * @param tensor Pointer to the tensor information. - * @param src Pointer to the source data in Q4.0 format. - * @param dst Pointer to the destination buffer where transformed data will be - * stored. - */ -GGML_CALL static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor, - const void* src, - void* dst) { - - int64_t n_elems = ggml_nelements(tensor); - int64_t groups = n_elems / QK4_0; - size_t quant_bytes = n_elems * sizeof(uint8_t) / 2; - - uint8_t* quant_offset = (uint8_t*)dst; - uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes); - - for (int i = 0; i < groups; i++) { - const block_q4_0* group = - (const block_q4_0*)((const char*)src + i * sizeof(block_q4_0)); - *scale_offset = group->d; - scale_offset++; - - // 0-15 - for (int j = 0; j < QK4_0 / 2; j += 2) { - (*quant_offset) = (group->qs[j] & 0x0F); - (*quant_offset) |= ((group->qs[j + 1] << 4)); - quant_offset++; - } - - // 16-31 - for (int j = 0; j < QK4_0 / 2; j += 2) { - (*quant_offset) = (group->qs[j] >> 4); - (*quant_offset) |= (group->qs[j + 1] & 0xF0); - quant_offset++; - } - } - - // put (uint4b_t -8) into int4b_t - for (quant_offset = (uint8_t*)dst; - quant_offset < (uint8_t*)dst + quant_bytes; quant_offset++) { - (*quant_offset) ^= 0x88; - } -} - -/** - * @brief Transform CANN processed data back into quantized Q4.0 format. - * - * This function transforms CANN processed data back into quantized Q4.0 format. - * It reverses the transformation performed by - * ggml_backend_cann_transform_q4_0(), converting the data back into its - * original quantized form. - * - * @param tensor Pointer to the tensor information. - * @param src Pointer to the source buffer containing transformed data. - * @param dst Pointer to the destination buffer where the Q4.0 formatted data - * will be stored. - */ -GGML_CALL static void ggml_backend_cann_transform_back_q4_0( - const ggml_tensor* tensor, void* src, void* dst) { - - int64_t n_elems = ggml_nelements(tensor); - int64_t groups = n_elems / QK4_0; - size_t quant_bytes = n_elems * sizeof(uint8_t) / 2; - - uint8_t* quant_offset = (uint8_t*)src; - uint16_t* scale_offset = (uint16_t*)((char*)src + quant_bytes); - - for (; quant_offset < (uint8_t*)src + quant_bytes; quant_offset++) { - (*quant_offset) ^= 0x88; - } - quant_offset = (uint8_t*)src; - - for (int i = 0; i < groups; i++) { - block_q4_0* group = (block_q4_0*)((char*)dst + i * sizeof(block_q4_0)); - group->d = *scale_offset; - scale_offset++; - - // 0-15 - for (int j = 0; j < QK4_0 / 2; j += 2) { - group->qs[j] = ((*quant_offset) & 0x0F); - group->qs[j + 1] = ((*quant_offset) >> 4); - quant_offset++; - } - - // 16-31 - for (int j = 0; j < QK4_0 / 2; j += 2) { - group->qs[j] |= ((*quant_offset) << 4); - group->qs[j + 1] |= ((*quant_offset) & 0xF0); - quant_offset++; - } - } -} - -/** - * @brief Transform quantized Q8.0 tensor data into a format suitable for CANN - * processing. - * - * This function transforms quantized Q8.0 tensor data into a format suitable - * for CANN processing. It extracts quantization values and scales from the - * source data and prepares them in a format expected by CANN operations. - * - * @param tensor Pointer to the tensor information. - * @param src Pointer to the source data in Q8.0 format. - * @param dst Pointer to the destination buffer where transformed data will be - * stored. - */ -GGML_CALL static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor, - const void* src, - void* dst) { - int64_t n_elems = ggml_nelements(tensor); - int64_t groups = n_elems / QK8_0; - size_t quant_bytes = n_elems * sizeof(uint8_t); - - uint8_t* quant_offset = (uint8_t*)dst; - uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes); - - for (int i = 0; i < groups; i++) { - const block_q8_0* group = - (const block_q8_0*)((const char*)src + i * sizeof(block_q8_0)); - *scale_offset = group->d; - scale_offset++; - size_t group_quant_size = QK8_0 * sizeof(uint8_t); - memcpy(quant_offset, group->qs, group_quant_size); - quant_offset += group_quant_size; - } -} - -/** - * @brief Transform CANN processed data back into quantized Q8.0 format. - * - * This function transforms CANN processed data back into quantized Q8.0 format. - * It reverses the transformation performed by - * ggml_backend_cann_transform_q8_0(), converting the data back into its - * original quantized form. - * - * @param tensor Pointer to the tensor information. - * @param src Pointer to the source buffer containing transformed data. - * @param dst Pointer to the destination buffer where the Q8.0 formatted data - * will be stored. - */ -GGML_CALL static void ggml_backend_cann_transform_back_q8_0( - const ggml_tensor* tensor, const void* src, void* dst) { - int64_t n_elems = ggml_nelements(tensor); - int64_t groups = n_elems / QK8_0; - size_t quant_bytes = n_elems * sizeof(uint8_t); - - const uint8_t* quant_offset = (const uint8_t*)src; - const uint16_t* scale_offset = - (const uint16_t*)((const char*)src + quant_bytes); - - for (int i = 0; i < groups; i++) { - block_q8_0* group = (block_q8_0*)((char*)dst + i * sizeof(block_q8_0)); - group->d = *scale_offset; - scale_offset++; - size_t group_quant_size = QK8_0 * sizeof(uint8_t); - memcpy(group->qs, quant_offset, group_quant_size); - quant_offset += group_quant_size; - } -} - -/** - * @brief Transform tensor data based on its type for CANN processing. - * - * This function transforms tensor data based on its quantization type for CANN - * processing. It dispatches the transformation based on the tensor's type to - * specialized functions handling Q4.0 and Q8.0 formats. - * - * @param tensor Pointer to the tensor information. - * @param src Pointer to the source data to be transformed. - * @param dst Pointer to the destination buffer where transformed data will be - * stored. - */ -GGML_CALL static void ggml_backend_cann_transform(ggml_tensor* tensor, - const void* src, void* dst) { - switch (tensor->type) { - case GGML_TYPE_Q4_0: - ggml_backend_cann_transform_q4_0(tensor, src, dst); - break; - case GGML_TYPE_Q8_0: - ggml_backend_cann_transform_q8_0(tensor, src, dst); - break; - default: - break; - } -} - -/** - * @brief Transform CANN processed data back into tensor data based on its type. - * - * This function transforms CANN processed data back into tensor data based on - * its quantization type for Q4.0 and Q8.0 formats. It dispatches the - * transformation based on the tensor's type to specialized functions. - * - * @param tensor Pointer to the tensor information. - * @param src Pointer to the source data containing CANN processed data. - * @param dst Pointer to the destination buffer where transformed tensor data - * will be stored. - */ -GGML_CALL static void ggml_backend_cann_transform_back( - const ggml_tensor* tensor, void* src, void* dst) { - switch (tensor->type) { - case GGML_TYPE_Q4_0: - ggml_backend_cann_transform_back_q4_0(tensor, src, dst); - break; - case GGML_TYPE_Q8_0: - ggml_backend_cann_transform_back_q8_0(tensor, src, dst); - break; - default: - break; - } -} - -/** - * @brief Check if transformation is needed for a given tensor type. - * - * This function checks if transformation is needed for a given tensor type - * to prepare data for CANN processing. - * - * @param type The tensor type to check. - * @return true if transformation is needed, false otherwise. - */ -GGML_CALL static bool need_transform(ggml_type type) { - switch (type) { - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q8_0: - return true; - default: - return false; - } -} - -/** - * @brief Initialize a tensor using data from a CANN buffer. - * - * This function initializes a tensor using data from a CANN buffer. - * It handles special cases such as views and quantization. - * - * @param buffer The CANN buffer from which to initialize the tensor. - * @param tensor Pointer to the tensor to be initialized. - */ -GGML_CALL static void ggml_backend_cann_buffer_init_tensor( - ggml_backend_buffer_t buffer, ggml_tensor* tensor) { - if (tensor->view_src != NULL && tensor->view_offs == 0) { - GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft); - return; - } - - // TODO: can backend doesn't support quantized yet. Just leave the code - // here. - if (ggml_is_quantized(tensor->type)) { - // Initialize padding to 0 to avoid possible NaN values - size_t original_size = ggml_nbytes(tensor); - size_t padded_size = - ggml_backend_buft_get_alloc_size(buffer->buft, tensor); - - if (padded_size > original_size && tensor->view_src == nullptr) { - size_t memset_size = padded_size - original_size; - ACL_CHECK(aclrtMemset((char*)tensor->data + original_size, - memset_size, 0, memset_size)); - } - } -} - -// TODO: need handle tensor which has paddings. -/** - * @brief Set tensor data in a CANN buffer. - * - * This function sets tensor data in a CANN buffer, handling transformations - * if needed based on the tensor's type. - * - * @param buffer The CANN buffer where the tensor data will be set. - * @param tensor Pointer to the tensor whose data will be set. - * @param data Pointer to the source data to be copied into the tensor. - * @param offset Offset in the source data from where to start copying. - * @param size Size of the data to be copied, in bytes. - */ -GGML_CALL static void ggml_backend_cann_buffer_set_tensor( - ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data, - size_t offset, size_t size) { - ggml_backend_cann_buffer_context *ctx = - (ggml_backend_cann_buffer_context *)buffer->context; - - ggml_cann_set_device(ctx->device); - // TODO: refer to cann(#6017), it use thread's default stream. - // For acl, synchronous functions use this default stream. - // Why aclrtSynchronizeDevice? - - if (!need_transform(tensor->type)) { - ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size, - ACL_MEMCPY_HOST_TO_DEVICE)); - } else { - void *transform_buffer = malloc(size); - ggml_backend_cann_transform(tensor, data, transform_buffer); - -#ifndef NDEBUG - void *check_buffer = malloc(size); - ggml_backend_cann_transform_back(tensor, transform_buffer, - check_buffer); - GGML_ASSERT(memcmp(data, check_buffer, size) == 0); - free(check_buffer); -#endif - ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, - transform_buffer, size, - ACL_MEMCPY_HOST_TO_DEVICE)); - free(transform_buffer); - } -} - -/** - * @brief Get tensor data from a CANN buffer. - * - * This function retrieves tensor data from a CANN buffer, handling - * transformations if needed based on the tensor's type. - * - * @param buffer The CANN buffer from which to retrieve tensor data. - * @param tensor Pointer to the tensor whose data will be retrieved. - * @param data Pointer to the destination buffer where the tensor data will be - * copied. - * @param offset Offset in the destination buffer where to start copying. - * @param size Size of the data to be copied, in bytes. - */ -GGML_CALL static void ggml_backend_cann_buffer_get_tensor( - ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data, - size_t offset, size_t size) { - ggml_backend_cann_buffer_context* ctx = - (ggml_backend_cann_buffer_context*)buffer->context; - - ggml_cann_set_device(ctx->device); - - if (!need_transform(tensor->type)) { - ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size, - ACL_MEMCPY_DEVICE_TO_HOST)); - } else { - void* transform_buffer = malloc(size); - ACL_CHECK(aclrtMemcpy(transform_buffer, size, - (char*)tensor->data + offset, size, - ACL_MEMCPY_DEVICE_TO_HOST)); - ggml_backend_cann_transform_back(tensor, transform_buffer, data); - free(transform_buffer); - } -} - -/** - * @brief Copy tensor data between CANN buffers if possible. - * - * This function copies tensor data between CANN buffers if the source and - * destination buffers are CANN buffers and they meet the necessary conditions - * (same device or devices can access each other). - * - * @param buffer The destination CANN buffer where the tensor data will be - * copied. - * @param src Pointer to the source tensor whose data will be copied. - * @param dst Pointer to the destination tensor where the data will be copied. - * @return true if the copy operation succeeded, false otherwise. - */ -GGML_CALL static bool ggml_backend_cann_buffer_cpy_tensor( - ggml_backend_buffer_t buffer, const ggml_tensor* src, ggml_tensor* dst) { - if (ggml_backend_buffer_is_cann(src->buffer)) { - ggml_backend_cann_buffer_context* src_ctx = - (ggml_backend_cann_buffer_context*)src->buffer->context; - ggml_backend_cann_buffer_context* dst_ctx = - (ggml_backend_cann_buffer_context*)buffer->context; - - size_t memcpy_size = ggml_nbytes(src); - // Same device. - if (src_ctx->device == dst_ctx->device) { - ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size, - (const char*)src->data, memcpy_size, - ACL_MEMCPY_DEVICE_TO_DEVICE)); - return true; - } else { - // Different device but can access by peer. - int32_t canAccessPeer = 0; - ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, src_ctx->device, - dst_ctx->device)); - if (canAccessPeer) { - ggml_cann_set_device(src_ctx->device); - ACL_CHECK(aclrtDeviceEnablePeerAccess(dst_ctx->device, 0)); - ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size, - (const char*)src->data, memcpy_size, - ACL_MEMCPY_DEVICE_TO_DEVICE)); - return true; - } - } - } - return false; -} - -/** - * @brief Clear a CANN buffer by setting all its memory to a specified value. - * - * This function clears a CANN buffer by setting all its memory to a specified - * value. - * - * @param buffer The CANN buffer to be cleared. - * @param value The value to which each byte in the buffer will be set. - */ -GGML_CALL static void ggml_backend_cann_buffer_clear( - ggml_backend_buffer_t buffer, uint8_t value) { - ggml_backend_cann_buffer_context* ctx = - (ggml_backend_cann_buffer_context*)buffer->context; - - ggml_cann_set_device(ctx->device); - ACL_CHECK(aclrtMemset(ctx->dev_ptr, buffer->size, value, buffer->size)); -} - -/** - * @brief Interface for a CANN buffer in the backend. - * - * This structure defines function pointers to operations that can be performed - * on a CANN buffer within the backend. - */ -static ggml_backend_buffer_i ggml_backend_cann_buffer_interface = { - /* .get_name = */ ggml_backend_cann_buffer_get_name, - /* .free_buffer = */ ggml_backend_cann_buffer_free_buffer, - /* .get_base = */ ggml_backend_cann_buffer_get_base, - /* .init_tensor = */ ggml_backend_cann_buffer_init_tensor, - /* .set_tensor = */ ggml_backend_cann_buffer_set_tensor, - /* .get_tensor = */ ggml_backend_cann_buffer_get_tensor, - /* .cpy_tensor = */ ggml_backend_cann_buffer_cpy_tensor, - /* .clear = */ ggml_backend_cann_buffer_clear, - /* .reset = */ NULL, -}; - -// cann buffer type -/** - * @brief Structure representing context information for a specific backend - * buffer type. - */ -struct ggml_backend_cann_buffer_type_context { - int32_t - device; /**< Device identifier associated with the buffer context. */ - std::string name; /**< Name associated with the buffer context. */ -}; - -/** - * @brief Retrieves the name associated with a CANN buffer type. - * - * This function returns the descriptive name associated with the specified - * CANN buffer type context. - * - * @param buft Pointer to the buffer type context. - * @return Const pointer to the C-style string containing the name. - */ -GGML_CALL static const char* ggml_backend_cann_buffer_type_name( - ggml_backend_buffer_type_t buft) { - return "CANN"; - - GGML_UNUSED(buft); -} - -/** - * @brief Allocates a new CANN buffer of the specified type and size. - * - * This function allocates a new CANN buffer on the specified device with the - * given size. - * - * @param buft Pointer to the buffer type context. - * @param size Size in bytes of the buffer to allocate. - * @return Pointer to the allocated buffer, or nullptr if allocation fails. - */ -GGML_CALL static ggml_backend_buffer_t -ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, - size_t size) { - ggml_backend_cann_buffer_type_context* buft_ctx = - (ggml_backend_cann_buffer_type_context*)buft->context; - - ggml_cann_set_device(buft_ctx->device); - - size = std::max(size, (size_t)1); - - void* dev_ptr; - aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST); - if (err != ACL_SUCCESS) { - GGML_CANN_LOG_ERROR( - "%s: allocating %.2f MiB on device %d: aclrtMalloc failed: %s\n", - __func__, size / 1024.0 / 1024.0, buft_ctx->device, - aclGetRecentErrMsg()); - return nullptr; - } - - ggml_backend_cann_buffer_context* ctx = - new ggml_backend_cann_buffer_context(buft_ctx->device, dev_ptr); - - return ggml_backend_buffer_init(buft, ggml_backend_cann_buffer_interface, - ctx, size); -} - -/** - * @brief Retrieves the memory alignment requirement for CANN buffers of this - * type. - * - * This function returns the alignment requirement in bytes for memory allocated - * by the CANN buffer type. - * - * @param buft Pointer to the buffer type context (unused in this - * implementation). - * @return The alignment requirement in bytes (fixed at 128 bytes for CANN - * buffers). - */ -GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alignment( - ggml_backend_buffer_type_t buft) { - return 128; - - GGML_UNUSED(buft); -} - -/** - * @brief Calculates the allocation size required for a tensor in a CANN buffer. - * - * Computes the total allocation size needed for storing the tensor's data in a - * CANN buffer, considering any necessary padding or adjustments for quantized - * types. - * - * @param buft Pointer to the buffer type context (unused in this - * implementation). - * @param tensor Pointer to the tensor for which the allocation size is - * calculated. - * @return The total allocation size in bytes required for the tensor in the - * CANN buffer. - */ -GGML_CALL static size_t ggml_backend_cann_buffer_type_get_alloc_size( - ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) { - size_t size = ggml_nbytes(tensor); - int64_t ne0 = tensor->ne[0]; - - // last line must bigger than 32, because every single op deal at - // least 32 bytes. - // TODO: quantized type? - // int64_t line_size = ne0 * ggml_element_size(tensor); - // int64_t line_size_align_32 = (line_size + 31) & ~31; - // size += (line_size_align_32 - line_size); - - // TODO: not support quantized yet. - // TODO: consider un-continue tensor. - if (ggml_is_quantized(tensor->type)) { - if (ne0 % MATRIX_ROW_PADDING != 0) { - size += ggml_row_size( - tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING); - } - } - - return size; - - GGML_UNUSED(buft); -} - -/** - * @brief Interface for managing CANN buffer types in the GGML backend. - * - * Provides function pointers for allocating, querying properties, and managing - * memory for CANN buffer types in the GGML backend. - */ -static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = { - /* .get_name = */ ggml_backend_cann_buffer_type_name, - /* .alloc_buffer = */ ggml_backend_cann_buffer_type_alloc_buffer, - /* .get_alignment = */ ggml_backend_cann_buffer_type_get_alignment, - /* .get_max_size = */ NULL, // defaults to SIZE_MAX - /* .get_alloc_size = */ ggml_backend_cann_buffer_type_get_alloc_size, - /* .is_host = */ NULL, -}; - -/** - * @brief Retrieves the CANN buffer type for a specified device. - * - * This function initializes and returns the buffer type interface associated - * with the given device. It ensures thread-safe access using a mutex. - * - * @param device The device index for which to retrieve the buffer type. - * @return A pointer to the buffer type interface for the specified device, or - * nullptr if the device index is out of range. - */ -GGML_CALL ggml_backend_buffer_type_t -ggml_backend_cann_buffer_type(int32_t device) { - static std::mutex mutex; - std::lock_guard lock(mutex); - - if (device >= ggml_backend_cann_get_device_count()) { - return nullptr; - } - - static ggml_backend_buffer_type - ggml_backend_cann_buffer_types[GGML_CANN_MAX_DEVICES]; - - static bool ggml_backend_cann_buffer_type_initialized = false; - - if (!ggml_backend_cann_buffer_type_initialized) { - for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) { - ggml_backend_cann_buffer_types[i] = { - /* .iface = */ ggml_backend_cann_buffer_type_interface, - /* .context = */ - new ggml_backend_cann_buffer_type_context{ - i, "CANN" + std::to_string(i)}, - }; - } - ggml_backend_cann_buffer_type_initialized = true; - } - - return &ggml_backend_cann_buffer_types[device]; -} - -/** - * @brief Computes the forward operation for a given tensor using CANN - * operations. - * - * This function selects the appropriate CANN operation based on the type of - * operation specified in the tensor and performs the computation. - * - * @param ctx The CANN context containing necessary resources and - * configurations. - * @param dst The destination tensor where the result of the computation will be - * stored. - * @return true if the computation was successful; false otherwise. - */ -static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx, - struct ggml_tensor* dst) { - switch (dst->op) { - case GGML_OP_REPEAT: - ggml_cann_repeat(ctx, dst); - break; - case GGML_OP_GET_ROWS: - ggml_cann_get_rows(ctx, dst); - break; - case GGML_OP_DUP: - ggml_cann_dup(ctx, dst); - break; - case GGML_OP_ADD: - ggml_cann_add(ctx, dst); - break; - case GGML_OP_ACC: - ggml_cann_acc(ctx, dst); - break; - case GGML_OP_MUL: - ggml_cann_mul_div(ctx, dst); - break; - case GGML_OP_DIV: - ggml_cann_mul_div(ctx, dst); - break; - case GGML_OP_UNARY: - switch (ggml_get_unary_op(dst)) { - case GGML_UNARY_OP_GELU: - ggml_cann_activation( - ctx, dst); - break; - case GGML_UNARY_OP_SILU: - ggml_cann_activation( - ctx, dst); - break; - // TODO: Use faster gelu?? - case GGML_UNARY_OP_GELU_QUICK: - ggml_cann_activation( - ctx, dst); - break; - case GGML_UNARY_OP_TANH: - ggml_cann_activation( - ctx, dst); - break; - case GGML_UNARY_OP_RELU: - ggml_cann_activation( - ctx, dst); - break; - case GGML_UNARY_OP_HARDSIGMOID: - ggml_cann_activation(ctx, dst); - break; - case GGML_UNARY_OP_HARDSWISH: - ggml_cann_activation(ctx, dst); - break; - default: - return false; - } - break; - case GGML_OP_NORM: - ggml_cann_norm(ctx, dst); - break; - case GGML_OP_GROUP_NORM: - ggml_cann_group_norm(ctx, dst); - break; - case GGML_OP_CONCAT: - ggml_cann_concat(ctx, dst); - break; - case GGML_OP_UPSCALE: - ggml_cann_upsample_nearest2d(ctx, dst); - break; - case GGML_OP_PAD: - ggml_cann_pad(ctx, dst); - break; - case GGML_OP_ARANGE: - ggml_cann_arange(ctx, dst); - break; - case GGML_OP_TIMESTEP_EMBEDDING: - ggml_cann_timestep_embedding(ctx, dst); - break; - case GGML_OP_LEAKY_RELU: - ggml_cann_leaky_relu(ctx, dst); - break; - case GGML_OP_RMS_NORM: - ggml_cann_rms_norm(ctx, dst); - break; - case GGML_OP_MUL_MAT: - ggml_cann_mul_mat(ctx, dst); - break; - case GGML_OP_MUL_MAT_ID: - return false; - case GGML_OP_SCALE: - ggml_cann_scale(ctx, dst); - break; - case GGML_OP_SQR: - ggml_cann_sqr(ctx, dst); - break; - case GGML_OP_CLAMP: - ggml_cann_clamp(ctx, dst); - break; - case GGML_OP_CPY: - ggml_cann_cpy(ctx, dst); - break; - case GGML_OP_CONT: - ggml_cann_dup(ctx, dst); - break; - case GGML_OP_NONE: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_PERMUTE: - case GGML_OP_TRANSPOSE: - break; - case GGML_OP_DIAG_MASK_INF: - ggml_cann_diag_mask(ctx, dst, -INFINITY); - break; - case GGML_OP_SOFT_MAX: - ggml_cann_softmax(ctx, dst); - break; - case GGML_OP_ROPE: - ggml_cann_rope(ctx, dst); - break; - case GGML_OP_IM2COL: - ggml_cann_im2col(ctx, dst); - break; - case GGML_OP_POOL_2D: - ggml_cann_pool2d(ctx, dst); - break; - case GGML_OP_SUM_ROWS: - ggml_cann_sum_rows(ctx, dst); - break; - case GGML_OP_ARGSORT: - ggml_cann_argsort(ctx, dst); - break; - default: - return false; - } - - return true; -} - -// backend -/** - * @brief Retrieves the name associated with the CANN backend. - * - * This function returns the name assigned to the CANN backend, which is stored - * in the context of the provided backend structure. - * - * @param backend Pointer to the CANN backend structure. - * @return A pointer to a constant string representing the backend name. - */ -GGML_CALL static const char* ggml_backend_cann_name(ggml_backend_t backend) { - ggml_backend_cann_context* cann_ctx = - (ggml_backend_cann_context*)backend->context; - - return cann_ctx->name.c_str(); -} - -/** - * @brief Frees resources associated with the CANN backend. - * - * This function releases resources associated with the CANN backend context - * and resets the device associated with the backend to its initial state. - * - * @param backend Pointer to the CANN backend structure to be freed. - */ -GGML_CALL static void ggml_backend_cann_free(ggml_backend_t backend) { - ggml_backend_cann_context* cann_ctx = - (ggml_backend_cann_context*)backend->context; - ACL_CHECK(aclrtSynchronizeDevice()); - ACL_CHECK(aclrtResetDevice(cann_ctx->device)); - - // finalize when last backend freed. - if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) { - ACL_CHECK(aclFinalize()); - } - - delete cann_ctx; - delete backend; -} - -/** - * @brief Retrieves the default buffer type associated with the CANN backend. - * - * This function returns the buffer type specific to the device associated - * with the CANN backend. It is used to allocate buffers for computations - * performed by the backend. - * - * @param backend Pointer to the CANN backend structure. - * @return Pointer to the buffer type structure for the CANN backend. - */ -GGML_CALL static ggml_backend_buffer_type_t -ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) { - ggml_backend_cann_context* cann_ctx = - (ggml_backend_cann_context*)backend->context; - - return ggml_backend_cann_buffer_type(cann_ctx->device); -} - -/** - * @brief Sets tensor data asynchronously in the CANN backend. - * - * This function asynchronously sets tensor data in the CANN backend. Depending - * on the tensor type, it may perform data transformations before copying data - * to the device. - * - * @param backend Pointer to the CANN backend structure. - * @param tensor Pointer to the tensor structure to set data for. - * @param data Pointer to the host data to copy to the tensor. - * @param offset Offset in bytes within the host data. - * @param size Size of the data to copy in bytes. - */ -GGML_CALL static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend, - ggml_tensor *tensor, - const void *data, - size_t offset, - size_t size) { - ggml_backend_cann_context *cann_ctx = - (ggml_backend_cann_context *)backend->context; - - if (!need_transform(tensor->type)) { - ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data, - size, ACL_MEMCPY_HOST_TO_DEVICE, - cann_ctx->stream())); - } else { - void *transform_buffer = malloc(size); - ggml_backend_cann_transform(tensor, data, transform_buffer); - -#ifndef NDEBUG - void *check_buffer = malloc(size); - ggml_backend_cann_transform_back(tensor, transform_buffer, - check_buffer); - GGML_ASSERT(memcmp(data, check_buffer, size)); - free(check_buffer); -#endif - ACL_CHECK(aclrtMemcpyAsync( - (char *)tensor->data + offset, size, transform_buffer, size, - ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream())); - ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream())); - free(transform_buffer); - } -} - -GGML_CALL static void ggml_backend_cann_get_tensor_async( - ggml_backend_t backend, const ggml_tensor *tensor, void *data, - size_t offset, size_t size) { - ggml_backend_cann_context *cann_ctx = - (ggml_backend_cann_context *)backend->context; - ggml_backend_buffer_t buf = - tensor->view_src ? tensor->view_src->buffer : tensor->buffer; - - GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) && - "unsupported buffer type"); - - if (!need_transform(tensor->type)) { - ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset, - size, ACL_MEMCPY_DEVICE_TO_HOST, - cann_ctx->stream())); - } else { - void *transform_buffer = malloc(size); - ACL_CHECK(aclrtMemcpyAsync( - transform_buffer, size, (char *)tensor->data + offset, size, - ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream())); - ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream())); - ggml_backend_cann_transform_back(tensor, transform_buffer, data); - free(transform_buffer); - } -} - -/** - * @brief Asynchronously copies tensor data between CANN backends. - * - * This function copies tensor data asynchronously between two CANN backends. It - * checks if both tensors reside in CANN buffers and whether the devices support - * peer-to-peer access for direct copying. If not, it returns false. - * - * @param backend_src Pointer to the source CANN backend structure. - * @param backend_dst Pointer to the destination CANN backend structure. - * @param src Pointer to the source tensor to copy data from. - * @param dst Pointer to the destination tensor to copy data to. - * @return true if the copy operation succeeds, false otherwise. - */ -GGML_CALL static bool ggml_backend_cann_cpy_tensor_async( - ggml_backend_t backend_src, ggml_backend_t backend_dst, - const ggml_tensor* src, ggml_tensor* dst) { - GGML_ASSERT(ggml_backend_is_cann(backend_src) || - ggml_backend_is_cann(backend_dst)); - - if (!ggml_backend_buffer_is_cann(src->buffer) || - !ggml_backend_buffer_is_cann(dst->buffer)) { - return false; - } - - ggml_backend_buffer_t buf_src = - src->view_src ? src->view_src->buffer : src->buffer; - ggml_backend_buffer_t buf_dst = - dst->view_src ? dst->view_src->buffer : dst->buffer; - - ggml_backend_cann_context* cann_ctx_src = - (ggml_backend_cann_context*)backend_src->context; - ggml_backend_cann_context* cann_ctx_dst = - (ggml_backend_cann_context*)backend_dst->context; - - size_t copy_size = ggml_nbytes(dst); - if (backend_src != backend_dst) { - ggml_backend_cann_buffer_context* buf_ctx_src = - (ggml_backend_cann_buffer_context*)buf_src->context; - ggml_backend_cann_buffer_context* buf_ctx_dst = - (ggml_backend_cann_buffer_context*)buf_dst->context; - - GGML_ASSERT(cann_ctx_src->device == buf_ctx_src->device); - GGML_ASSERT(cann_ctx_dst->device == buf_ctx_dst->device); - - int32_t canAccessPeer = 0; - ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, cann_ctx_src->device, - cann_ctx_dst->device)); - if (!canAccessPeer) { - return false; - } - - // need open both directions for memcpyasync between devices. - ggml_cann_set_device(cann_ctx_dst->device); - ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_src->device, 0)); - ggml_cann_set_device(cann_ctx_src->device); - ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0)); - - ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size, - ACL_MEMCPY_DEVICE_TO_DEVICE, - cann_ctx_src->stream())); - - //TODO: workaround for Event didn`t work here. - aclrtSynchronizeStream(cann_ctx_src->stream()); - } else { - // src and dst are on the same backend - ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size, - ACL_MEMCPY_DEVICE_TO_DEVICE, - cann_ctx_dst->stream())); - } - - return true; -} - -/** - * @brief Synchronizes a CANN backend. - * - * This function synchronizes the specified CANN backend by waiting for all - * operations in its associated stream to complete. - * - * @param backend Pointer to the CANN backend structure to synchronize. - */ -GGML_CALL static void ggml_backend_cann_synchronize(ggml_backend_t backend) { - ggml_backend_cann_context* cann_ctx = - (ggml_backend_cann_context*)backend->context; - - ggml_cann_set_device(cann_ctx->device); - - ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream())); -} - -/** - * @brief Computes a computational graph using a CANN backend. - * - * This function computes the operations defined in the computational graph - * using the specified CANN backend. - * - * @param backend Pointer to the CANN backend structure to use for computation. - * @param cgraph Pointer to the computational graph structure containing nodes - * representing operations to be computed. - * @return enum ggml_status Returns GGML_STATUS_SUCCESS if computation - * completes successfully, otherwise an appropriate error status. - */ -GGML_CALL static enum ggml_status ggml_backend_cann_graph_compute( - ggml_backend_t backend, ggml_cgraph* cgraph) { - ggml_backend_cann_context* cann_ctx = - (ggml_backend_cann_context*)backend->context; - - ggml_cann_set_device(cann_ctx->device); - - for (int i = 0; i < cgraph->n_nodes; i++) { - ggml_tensor* node = cgraph->nodes[i]; - - if (ggml_is_empty(node) || node->op == GGML_OP_NONE) { - continue; - } - - bool ok = ggml_cann_compute_forward(*cann_ctx, node); - - if (!ok) { - GGML_CANN_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, - node->name, ggml_op_name(node->op)); - } - GGML_ASSERT(ok); - } - - return GGML_STATUS_SUCCESS; -} - -/** - * @brief Checks if the CANN backend supports a specific operation. - * - * This function checks whether the specified operation is supported by the - * CANN backend. - * - * @param backend Pointer to the CANN backend structure to check support for - * the operation. - * @param op Pointer to the tensor representing the operation to check. - * @return bool Returns true if the operation is supported by the backend, - * otherwise false. - */ -GGML_CALL static bool ggml_backend_cann_supports_op(ggml_backend_t backend, - const ggml_tensor* op) { - switch (op->op) { - case GGML_OP_UNARY: - switch (ggml_get_unary_op(op)) { - case GGML_UNARY_OP_GELU: - case GGML_UNARY_OP_SILU: - case GGML_UNARY_OP_RELU: - case GGML_UNARY_OP_HARDSIGMOID: - case GGML_UNARY_OP_HARDSWISH: - case GGML_UNARY_OP_GELU_QUICK: - case GGML_UNARY_OP_TANH: - return true; - default: - return false; - } - case GGML_OP_MUL_MAT: { - switch (op->src[0]->type) { - case GGML_TYPE_F16: - case GGML_TYPE_F32: - case GGML_TYPE_Q8_0: - // TODO: fix me - // Current groupsize should not be greater than k-1 in - // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(). - case GGML_TYPE_Q4_0: - return true; - default: - return false; - } - } - case GGML_OP_MUL_MAT_ID: - return false; - // embedding - case GGML_OP_GET_ROWS: { - switch (op->src[0]->type) { - case GGML_TYPE_F32: - case GGML_TYPE_F16: - case GGML_TYPE_Q4_0: - case GGML_TYPE_Q8_0: - return true; - default: - return false; - } - } break; - case GGML_OP_CPY: { - switch (op->type) { - case GGML_TYPE_F32: - case GGML_TYPE_F16: - case GGML_TYPE_Q8_0: - case GGML_TYPE_Q4_0: - return true; - default: - return false; - } - } - case GGML_OP_DUP: - case GGML_OP_REPEAT: - case GGML_OP_CONCAT: - case GGML_OP_NONE: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_PERMUTE: - case GGML_OP_TRANSPOSE: - case GGML_OP_NORM: - case GGML_OP_ADD: - case GGML_OP_MUL: - case GGML_OP_DIV: - case GGML_OP_RMS_NORM: - case GGML_OP_SCALE: - case GGML_OP_SQR: - case GGML_OP_CLAMP: - case GGML_OP_CONT: - case GGML_OP_DIAG_MASK_INF: - case GGML_OP_SOFT_MAX: - case GGML_OP_ROPE: - case GGML_OP_IM2COL: - case GGML_OP_POOL_2D: - case GGML_OP_SUM_ROWS: - case GGML_OP_ARGSORT: - case GGML_OP_ACC: - case GGML_OP_GROUP_NORM: - case GGML_OP_UPSCALE: - case GGML_OP_PAD: - case GGML_OP_ARANGE: - case GGML_OP_TIMESTEP_EMBEDDING: - case GGML_OP_LEAKY_RELU: - return true; - default: - return false; - } - - GGML_UNUSED(backend); -} - -/** - * @brief Checks if the backend buffer type is associated with the CANN backend. - * - * This function checks whether the provided backend buffer type is associated - * with the CANN backend based on the comparison of its name retrieval function - * pointer. - * - * @param buft Pointer to the backend buffer type to check. - * @return bool Returns true if the buffer type is associated with the CANN - * backend, otherwise false. - */ -static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) { - return buft->iface.get_name == ggml_backend_cann_buffer_type_name; -} - -/** - * @brief Checks if the CANN backend supports a specific backend buffer type. - * - * This function determines whether the CANN backend supports the given backend - * buffer type by comparing the device context of the backend and buffer type. - * It returns true if the devices are same between the backend context and - * buffer type context. - * - * @param backend Pointer to the CANN backend. - * @param buft Pointer to the backend buffer type to check. - * @return bool Returns true if the CANN backend supports the buffer type, - * otherwise false. - */ -GGML_CALL static bool ggml_backend_cann_supports_buft( - ggml_backend_t backend, ggml_backend_buffer_type_t buft) { - if (ggml_backend_buft_is_cann(buft)) { - ggml_backend_cann_context * cann_ctx = - (ggml_backend_cann_context *)backend->context; - ggml_backend_cann_buffer_type_context * buft_ctx = - (ggml_backend_cann_buffer_type_context *)buft->context; - return buft_ctx->device == cann_ctx->device; - } - return false; -} - -/** - * @brief Determines if a tensor operation should be offloaded to the CANN - * backend. - * - * This function checks if a given tensor operation should be offloaded to the - * CANN backend based on the operation type and the size of the tensor. It - * returns true if the second dimension (ne[1]) of the tensor is greater than or - * equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS. - * - * @param backend Pointer to the CANN backend. - * @param op Pointer to the tensor operation to check. - * @return bool Returns true if the operation should be offloaded, otherwise - * false. - */ -GGML_CALL static bool ggml_backend_cann_offload_op(ggml_backend_t backend, - const ggml_tensor* op) { - const int min_batch_size = 32; - GGML_UNUSED(backend); - - return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS; -} - -/** - * @brief Creates a new event for the CANN backend. - * - * This function initializes a new event for the CANN backend by setting the - * device and creating an ACL runtime event. The created event is then wrapped - * in a ggml_backend_event structure and returned. - * - * @param backend Pointer to the CANN backend. - * @return ggml_backend_event_t Returns a pointer to the new event structure. - */ -static ggml_backend_event_t ggml_backend_cann_event_new( - ggml_backend_t backend) { - ggml_backend_cann_context* cann_ctx = - (ggml_backend_cann_context*)backend->context; - - ggml_cann_set_device(cann_ctx->device); - - aclrtEvent event; - ACL_CHECK(aclrtCreateEvent(&event)); - - return new ggml_backend_event{ - /* .backend = */ backend, - /* .context = */ event, - }; -} - -/** - * @brief Frees a CANN backend event. - * - * This function destroys the ACL runtime event associated with the given CANN - * backend event and then deletes the event structure itself. - * - * @param event Pointer to the event structure to be freed. - */ -static void ggml_backend_cann_event_free(ggml_backend_event_t event) { - ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context)); - - delete event; -} - -/** - * @brief Records an event on the CANN backend stream. - * - * This function records the given event on the ACL runtime stream associated - * with the backend context. - * - * @param event Pointer to the event structure to be recorded. - */ -static void ggml_backend_cann_event_record(ggml_backend_event_t event) { - ggml_backend_cann_context* cann_ctx = - (ggml_backend_cann_context*)event->backend->context; - - ACL_CHECK(aclrtRecordEvent((aclrtEvent)event->context, cann_ctx->stream())); -} - -/** - * @brief Waits for a recorded event to complete on the CANN backend stream. - * - * This function makes the given backend wait for the event to complete on its - * ACL runtime stream. - * - * @param backend Pointer to the backend structure. - * @param event Pointer to the event structure that the backend needs to wait - * for. - */ -static void ggml_backend_cann_event_wait(ggml_backend_t backend, - ggml_backend_event_t event) { - ggml_backend_cann_context* cann_ctx = - (ggml_backend_cann_context*)backend->context; - - if (ggml_backend_is_cann(event->backend)) { - ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(), - (aclrtEvent)event->context)); - } else { - GGML_ABORT("fatal error"); - } -} - -/** - * @brief Synchronizes the given event on the CANN backend. - * - * This function waits for the specified event to complete on the ACL runtime. - * - * @param event Pointer to the event structure to be synchronized. - */ -static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) { - ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context)); -} - -/** - * @brief Structure defining the interface for the CANN backend. - * - * This structure contains function pointers for various operations - * supported by the CANN backend, including name retrieval, memory - * management, tensor operations, synchronization, and event handling. - */ -static ggml_backend_i ggml_backend_cann_interface = { - /* .get_name = */ ggml_backend_cann_name, - /* .free = */ ggml_backend_cann_free, - /* .get_default_buffer_type = */ ggml_backend_cann_get_default_buffer_type, - /* .set_tensor_async = */ ggml_backend_cann_set_tensor_async, - /* .get_tensor_async = */ ggml_backend_cann_get_tensor_async, - /* .cpy_tensor_async = */ ggml_backend_cann_cpy_tensor_async, - /* .synchronize = */ ggml_backend_cann_synchronize, - /* .graph_plan_create = */ NULL, - /* .graph_plan_free = */ NULL, - /* .graph_plan_update = */ NULL, - /* .graph_plan_compute = */ NULL, - /* .graph_compute = */ ggml_backend_cann_graph_compute, - /* .supports_op = */ ggml_backend_cann_supports_op, - /* .supports_buft = */ ggml_backend_cann_supports_buft, - /* .offload_op = */ ggml_backend_cann_offload_op, - /* .event_new = */ ggml_backend_cann_event_new, - /* .event_free = */ ggml_backend_cann_event_free, - /* .event_record = */ ggml_backend_cann_event_record, - /* .event_wait = */ ggml_backend_cann_event_wait, - /* .event_synchronize = */ ggml_backend_cann_event_synchronize, -}; - -/** - * @brief Return the hardcoded GUID for the CANN backend. - * - * This function returns a static GUID which uniquely identifies the CANN - * backend. - * - * @return A pointer to the static GUID. - */ -static ggml_guid_t ggml_backend_cann_guid() { - static ggml_guid guid = {0xa1, 0x94, 0xaf, 0xac, 0xbd, 0x4f, 0x47, 0x34, - 0xbe, 0x1a, 0x9e, 0x71, 0x1f, 0x9e, 0xed, 0x64}; - return &guid; -} - -GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) { - aclInit(nullptr); - if (device < 0 || device >= ggml_backend_cann_get_device_count()) { - GGML_CANN_LOG_ERROR("%s: error: invalid device %d\n", __func__, device); - return nullptr; - } - - ggml_backend_cann_context* ctx = new ggml_backend_cann_context(device); - if (ctx == nullptr) { - GGML_CANN_LOG_ERROR("%s: error: failed to allocate context\n", __func__); - return nullptr; - } - - ggml_backend_t cann_backend = - new ggml_backend{/* .guid = */ ggml_backend_cann_guid(), - /* .interface = */ ggml_backend_cann_interface, - /* .context = */ ctx}; - - return cann_backend; -} - -GGML_CALL bool ggml_backend_is_cann(ggml_backend_t backend) { - return backend != NULL && - ggml_guid_matches(backend->guid, ggml_backend_cann_guid()); -} - -GGML_CALL int32_t ggml_backend_cann_get_device_count() { - return ggml_cann_info().device_count; -} - -GGML_CALL void ggml_backend_cann_get_device_description( - int32_t device, char* description, size_t description_size) { - ggml_cann_set_device(device); - const char* soc_name = aclrtGetSocName(); - snprintf(description, description_size, "%s", soc_name); -} - -GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device, size_t* free, - size_t* total) { - ggml_cann_set_device(device); - ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total)); -} - -// backend registry -/** - * @brief Initializes a CANN backend based on the provided parameters. - * - * This function initializes a CANN backend using the device index and then - * initializes the backend using `ggml_backend_cann_init`. - * - * @param params Parameters for initialization (unused in this implementation). - * @param user_data User data containing the device index to initialize the - * backend. - * @return ggml_backend_t The initialized CANN backend. - */ -GGML_CALL static ggml_backend_t ggml_backend_reg_cann_init(const char* params, - void* user_data) { - ggml_backend_t cann_backend = - ggml_backend_cann_init((int)(intptr_t)user_data); - return cann_backend; - - GGML_UNUSED(params); -} - -extern "C" GGML_CALL int ggml_backend_cann_reg_devices(); - -/** - * @brief Registers CANN (Ascend) devices as backend options. - * - * This function initializes ACL, retrieves the number of available CANN - * devices, and registers each device as a backend option using - * `ggml_backend_register`. Each device is given a unique name based on - * `GGML_CANN_NAME` followed by its index. - * - * @return int The number of CANN devices registered. - */ -GGML_CALL int ggml_backend_cann_reg_devices() { - uint32_t device_count = ggml_backend_cann_get_device_count(); - // initialization - for (uint32_t i = 0; i < device_count; i++) { - char name[128]; - snprintf(name, sizeof(name), "CANN%d", i); - ggml_backend_register(name, ggml_backend_reg_cann_init, - ggml_backend_cann_buffer_type(i), - (void*)(intptr_t)i); - } - return device_count; -}