From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mails.dpdk.org (mails.dpdk.org [217.70.189.124]) by inbox.dpdk.org (Postfix) with ESMTP id 04305456FF; Wed, 31 Jul 2024 08:42:17 +0200 (CEST) Received: from mails.dpdk.org (localhost [127.0.0.1]) by mails.dpdk.org (Postfix) with ESMTP id 72D454065D; Wed, 31 Jul 2024 08:42:17 +0200 (CEST) Received: from mx0b-0016f401.pphosted.com (mx0b-0016f401.pphosted.com [67.231.156.173]) by mails.dpdk.org (Postfix) with ESMTP id 1015D40652 for ; Wed, 31 Jul 2024 08:32:57 +0200 (CEST) Received: from pps.filterd (m0045851.ppops.net [127.0.0.1]) by mx0b-0016f401.pphosted.com (8.18.1.2/8.18.1.2) with ESMTP id 46V4fQnH026781; Tue, 30 Jul 2024 23:32:54 -0700 DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=marvell.com; h= cc:content-transfer-encoding:content-type:date:from:message-id :mime-version:subject:to; s=pfpt0220; bh=7rnHlV3zLQYN5rdXAkkDUiC Q+oJg7xmnv8+wtZ8WOmE=; b=Up0//2GYs81uxitN7LKkKk9+3ZWA21DFXbt4laP Ee0H4MhOqjr4QKIC0CQcDfxWAAx3SOSLJwdEM7i5P/UHY5TRJj5/kNbYpk2bG3gA 8U6oJBlsQqP0cpyhdCrVkKNjeyGPKerFUMEXXxFdNzynwhhdJMp8cqV2/G5EOiYx W49xFVzvdW3W/MrwEG94JNgqlYt8eDt6hPeVcnaIOHe64Te128aRwhSu1OKjaGv+ SZQ+FTfQfsrYHQKxZTRWgMTrT/3mgrhyysbVBJrbsQCiBuqHPhBvilBxt/B+pprl WnWi03dCkz01aJ1MoePWjjLy90r/yLbTXKiRXsaUMvp7F6Q== Received: from dc6wp-exch02.marvell.com ([4.21.29.225]) by mx0b-0016f401.pphosted.com (PPS) with ESMTPS id 40qeeurbe6-1 (version=TLSv1.2 cipher=ECDHE-RSA-AES256-GCM-SHA384 bits=256 verify=NOT); Tue, 30 Jul 2024 23:32:53 -0700 (PDT) Received: from DC6WP-EXCH02.marvell.com (10.76.176.209) by DC6WP-EXCH02.marvell.com (10.76.176.209) with Microsoft SMTP Server (version=TLS1_2, cipher=TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384) id 15.2.1544.4; Tue, 30 Jul 2024 23:32:52 -0700 Received: from maili.marvell.com (10.69.176.80) by DC6WP-EXCH02.marvell.com (10.76.176.209) with Microsoft SMTP Server id 15.2.1544.4 via Frontend Transport; Tue, 30 Jul 2024 23:32:52 -0700 Received: from ml-host-33.sclab.marvell.com. (unknown [10.110.143.233]) by maili.marvell.com (Postfix) with ESMTP id 535EF5C68E3; Tue, 30 Jul 2024 23:32:52 -0700 (PDT) From: Srikanth Yalavarthi To: Srikanth Yalavarthi , Wathsala Vithanage CC: , , , Subject: [PATCH v1 1/1] mldev: introduce data type conversion functions Date: Tue, 30 Jul 2024 23:32:49 -0700 Message-ID: <20240731063250.9008-1-syalavarthi@marvell.com> X-Mailer: git-send-email 2.45.1 MIME-Version: 1.0 Content-Transfer-Encoding: 8bit Content-Type: text/plain X-Proofpoint-GUID: 0GnPxi4TDeoU3-X-PnSeCJCYMhW3N1g_ X-Proofpoint-ORIG-GUID: 0GnPxi4TDeoU3-X-PnSeCJCYMhW3N1g_ X-Proofpoint-Virus-Version: vendor=baseguard engine=ICAP:2.0.293,Aquarius:18.0.1039,Hydra:6.0.680,FMLib:17.12.28.16 definitions=2024-07-31_03,2024-07-30_01,2024-05-17_01 X-BeenThere: dev@dpdk.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: DPDK patches and discussions List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: dev-bounces@dpdk.org Introduced data type conversion functions with support for user defined scale factor and zero-point. Updated library functions to support asymmetric / affine conversion for integer types. Signed-off-by: Srikanth Yalavarthi --- drivers/ml/cnxk/cnxk_ml_io.c | 134 +++---- lib/mldev/mldev_utils.h | 453 ----------------------- lib/mldev/mldev_utils_neon.c | 452 ++++++++++++++--------- lib/mldev/mldev_utils_neon_bfloat16.c | 20 +- lib/mldev/mldev_utils_scalar.c | 156 ++++---- lib/mldev/mldev_utils_scalar_bfloat16.c | 12 +- lib/mldev/rte_mldev.h | 462 ++++++++++++++++++++++++ lib/mldev/version.map | 40 +- 8 files changed, 936 insertions(+), 793 deletions(-) diff --git a/drivers/ml/cnxk/cnxk_ml_io.c b/drivers/ml/cnxk/cnxk_ml_io.c index 4b0adc2ae47..a418b7e684d 100644 --- a/drivers/ml/cnxk/cnxk_ml_io.c +++ b/drivers/ml/cnxk/cnxk_ml_io.c @@ -26,39 +26,40 @@ cnxk_ml_io_quantize_single(struct cnxk_ml_io *input, uint8_t *dbuffer, uint8_t * if (dtype == qtype) { rte_memcpy(qbuffer, dbuffer, input->sz_d); - } else { - switch (qtype) { - case RTE_ML_IO_TYPE_INT8: - ret = rte_ml_io_float32_to_int8(qscale, nb_elements, dbuffer, qbuffer); - break; - case RTE_ML_IO_TYPE_UINT8: - ret = rte_ml_io_float32_to_uint8(qscale, nb_elements, dbuffer, qbuffer); - break; - case RTE_ML_IO_TYPE_INT16: - ret = rte_ml_io_float32_to_int16(qscale, nb_elements, dbuffer, qbuffer); - break; - case RTE_ML_IO_TYPE_UINT16: - ret = rte_ml_io_float32_to_uint16(qscale, nb_elements, dbuffer, qbuffer); - break; - case RTE_ML_IO_TYPE_INT32: - ret = rte_ml_io_float32_to_int32(qscale, nb_elements, dbuffer, qbuffer); - break; - case RTE_ML_IO_TYPE_UINT32: - ret = rte_ml_io_float32_to_uint32(qscale, nb_elements, dbuffer, qbuffer); - break; - case RTE_ML_IO_TYPE_INT64: - ret = rte_ml_io_float32_to_int64(qscale, nb_elements, dbuffer, qbuffer); - break; - case RTE_ML_IO_TYPE_UINT64: - ret = rte_ml_io_float32_to_uint64(qscale, nb_elements, dbuffer, qbuffer); - break; - case RTE_ML_IO_TYPE_FP16: - ret = rte_ml_io_float32_to_float16(nb_elements, dbuffer, qbuffer); - break; - default: - plt_err("Unsupported qtype : %u", qtype); - ret = -ENOTSUP; - } + return ret; + } + + switch (qtype) { + case RTE_ML_IO_TYPE_INT8: + ret = rte_ml_io_float32_to_int8(dbuffer, qbuffer, nb_elements, 1.0 / qscale, 0); + break; + case RTE_ML_IO_TYPE_UINT8: + ret = rte_ml_io_float32_to_uint8(dbuffer, qbuffer, nb_elements, 1.0 / qscale, 0); + break; + case RTE_ML_IO_TYPE_INT16: + ret = rte_ml_io_float32_to_int16(dbuffer, qbuffer, nb_elements, 1.0 / qscale, 0); + break; + case RTE_ML_IO_TYPE_UINT16: + ret = rte_ml_io_float32_to_uint16(dbuffer, qbuffer, nb_elements, 1.0 / qscale, 0); + break; + case RTE_ML_IO_TYPE_INT32: + ret = rte_ml_io_float32_to_int32(dbuffer, qbuffer, nb_elements, 1.0 / qscale, 0); + break; + case RTE_ML_IO_TYPE_UINT32: + ret = rte_ml_io_float32_to_uint32(dbuffer, qbuffer, nb_elements, 1.0 / qscale, 0); + break; + case RTE_ML_IO_TYPE_INT64: + ret = rte_ml_io_float32_to_int64(dbuffer, qbuffer, nb_elements, 1.0 / qscale, 0); + break; + case RTE_ML_IO_TYPE_UINT64: + ret = rte_ml_io_float32_to_uint64(dbuffer, qbuffer, nb_elements, 1.0 / qscale, 0); + break; + case RTE_ML_IO_TYPE_FP16: + ret = rte_ml_io_float32_to_float16(dbuffer, qbuffer, nb_elements); + break; + default: + plt_err("Unsupported qtype : %u", qtype); + ret = -ENOTSUP; } return ret; @@ -80,39 +81,40 @@ cnxk_ml_io_dequantize_single(struct cnxk_ml_io *output, uint8_t *qbuffer, uint8_ if (dtype == qtype) { rte_memcpy(dbuffer, qbuffer, output->sz_q); - } else { - switch (qtype) { - case RTE_ML_IO_TYPE_INT8: - ret = rte_ml_io_int8_to_float32(dscale, nb_elements, qbuffer, dbuffer); - break; - case RTE_ML_IO_TYPE_UINT8: - ret = rte_ml_io_uint8_to_float32(dscale, nb_elements, qbuffer, dbuffer); - break; - case RTE_ML_IO_TYPE_INT16: - ret = rte_ml_io_int16_to_float32(dscale, nb_elements, qbuffer, dbuffer); - break; - case RTE_ML_IO_TYPE_UINT16: - ret = rte_ml_io_uint16_to_float32(dscale, nb_elements, qbuffer, dbuffer); - break; - case RTE_ML_IO_TYPE_INT32: - ret = rte_ml_io_int32_to_float32(dscale, nb_elements, qbuffer, dbuffer); - break; - case RTE_ML_IO_TYPE_UINT32: - ret = rte_ml_io_uint32_to_float32(dscale, nb_elements, qbuffer, dbuffer); - break; - case RTE_ML_IO_TYPE_INT64: - ret = rte_ml_io_int64_to_float32(dscale, nb_elements, qbuffer, dbuffer); - break; - case RTE_ML_IO_TYPE_UINT64: - ret = rte_ml_io_uint64_to_float32(dscale, nb_elements, qbuffer, dbuffer); - break; - case RTE_ML_IO_TYPE_FP16: - ret = rte_ml_io_float16_to_float32(nb_elements, qbuffer, dbuffer); - break; - default: - plt_err("Unsupported qtype: %u", qtype); - ret = -ENOTSUP; - } + return 0; + } + + switch (qtype) { + case RTE_ML_IO_TYPE_INT8: + ret = rte_ml_io_int8_to_float32(qbuffer, dbuffer, nb_elements, dscale, 0); + break; + case RTE_ML_IO_TYPE_UINT8: + ret = rte_ml_io_uint8_to_float32(qbuffer, dbuffer, nb_elements, dscale, 0); + break; + case RTE_ML_IO_TYPE_INT16: + ret = rte_ml_io_int16_to_float32(qbuffer, dbuffer, nb_elements, dscale, 0); + break; + case RTE_ML_IO_TYPE_UINT16: + ret = rte_ml_io_uint16_to_float32(qbuffer, dbuffer, nb_elements, dscale, 0); + break; + case RTE_ML_IO_TYPE_INT32: + ret = rte_ml_io_int32_to_float32(qbuffer, dbuffer, nb_elements, dscale, 0); + break; + case RTE_ML_IO_TYPE_UINT32: + ret = rte_ml_io_uint32_to_float32(qbuffer, dbuffer, nb_elements, dscale, 0); + break; + case RTE_ML_IO_TYPE_INT64: + ret = rte_ml_io_int64_to_float32(qbuffer, dbuffer, nb_elements, dscale, 0); + break; + case RTE_ML_IO_TYPE_UINT64: + ret = rte_ml_io_uint64_to_float32(qbuffer, dbuffer, nb_elements, dscale, 0); + break; + case RTE_ML_IO_TYPE_FP16: + ret = rte_ml_io_float16_to_float32(qbuffer, dbuffer, nb_elements); + break; + default: + plt_err("Unsupported qtype: %u", qtype); + ret = -ENOTSUP; } return ret; diff --git a/lib/mldev/mldev_utils.h b/lib/mldev/mldev_utils.h index 5e2a180adce..37c90b44a8e 100644 --- a/lib/mldev/mldev_utils.h +++ b/lib/mldev/mldev_utils.h @@ -52,459 +52,6 @@ __rte_internal void rte_ml_io_type_to_str(enum rte_ml_io_type type, char *str, int len); -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) to signed 8-bit - * integer format (INT8). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store INT8 numbers. Size of buffer is equal to (nb_elements * 1) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_int8(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in signed 8-bit integer format (INT8) to single precision - * floating format (float32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing INT8 numbers. Size of buffer is equal to (nb_elements * 1) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_int8_to_float32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) to unsigned - * 8-bit integer format (UINT8). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store UINT8 numbers. Size of buffer is equal to (nb_elements * 1) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_uint8(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in unsigned 8-bit integer format (UINT8) to single precision - * floating format (float32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing UINT8 numbers. Size of buffer is equal to (nb_elements * 1) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_uint8_to_float32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) to signed - * 16-bit integer format (INT16). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store INT16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_int16(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in signed 16-bit integer format (INT16) to single precision - * floating format (float32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing INT16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_int16_to_float32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) to unsigned - * 16-bit integer format (UINT16). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store UINT16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_uint16(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in unsigned 16-bit integer format (UINT16) to single - * precision floating format (float32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing UINT16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_uint16_to_float32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) - * to signed 32-bit integer format (INT32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store INT32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_int32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in signed 32-bit integer format (INT32) - * to single precision floating format (float32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing INT32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_int32_to_float32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) - * to unsigned 32-bit integer format (UINT32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store UINT32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_uint32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in unsigned 32-bit integer format (UINT32) - * to single precision floating format (float32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing UINT32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_uint32_to_float32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) - * to signed 64-bit integer format (INT64). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store INT64 numbers. Size of buffer is equal to (nb_elements * 8) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_int64(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in signed 64-bit integer format (INT64) - * to single precision floating format (float32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing INT64 numbers. Size of buffer is equal to (nb_elements * 8) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_int64_to_float32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) - * to unsigned 64-bit integer format (UINT64). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * @param[out] output - * Output buffer to store UINT64 numbers. Size of buffer is equal to (nb_elements * 8) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_uint64(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in unsigned 64-bit integer format (UINT64) - * to single precision floating format (float32). - * - * @param[in] scale - * Scale factor for conversion. - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing UINT64 numbers. Size of buffer is equal to (nb_elements * 8) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_uint64_to_float32(float scale, uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) to half - * precision floating point format (FP16). - * - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements *4) bytes. - * @param[out] output - * Output buffer to store float16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_float16(uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in half precision floating format (FP16) to single precision - * floating point format (float32). - * - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float16_to_float32(uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in single precision floating format (float32) to brain - * floating point format (bfloat16). - * - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements *4) bytes. - * @param[out] output - * Output buffer to store bfloat16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_float32_to_bfloat16(uint64_t nb_elements, void *input, void *output); - -/** - * @internal - * - * Convert a buffer containing numbers in brain floating point format (bfloat16) to single precision - * floating point format (float32). - * - * @param[in] nb_elements - * Number of elements in the buffer. - * @param[in] input - * Input buffer containing bfloat16 numbers. Size of buffer is equal to (nb_elements * 2) - * bytes. - * @param[out] output - * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. - * - * @return - * - 0, Success. - * - < 0, Error code on failure. - */ -__rte_internal -int -rte_ml_io_bfloat16_to_float32(uint64_t nb_elements, void *input, void *output); - #ifdef __cplusplus } #endif diff --git a/lib/mldev/mldev_utils_neon.c b/lib/mldev/mldev_utils_neon.c index 4cde2ebabd3..8751a40863e 100644 --- a/lib/mldev/mldev_utils_neon.c +++ b/lib/mldev/mldev_utils_neon.c @@ -17,7 +17,7 @@ */ static inline void -__float32_to_int8_neon_s8x8(float scale, float *input, int8_t *output) +__float32_to_int8_neon_s8x8(const float *input, int8_t *output, float scale, int8_t zero_point) { int16x4_t s16x4_l; int16x4_t s16x4_h; @@ -30,7 +30,8 @@ __float32_to_int8_neon_s8x8(float scale, float *input, int8_t *output) * Use round to nearest with ties away rounding mode. */ f32x4 = vld1q_f32(input); - f32x4 = vmulq_n_f32(f32x4, scale); + f32x4 = vdivq_f32(f32x4, vdupq_n_f32(scale)); + f32x4 = vaddq_f32(f32x4, vdupq_n_f32((float)zero_point)); s32x4 = vcvtaq_s32_f32(f32x4); s16x4_l = vqmovn_s32(s32x4); @@ -38,7 +39,8 @@ __float32_to_int8_neon_s8x8(float scale, float *input, int8_t *output) * Use round to nearest with ties away rounding mode. */ f32x4 = vld1q_f32(input + 4); - f32x4 = vmulq_n_f32(f32x4, scale); + f32x4 = vdivq_f32(f32x4, vdupq_n_f32(scale)); + f32x4 = vaddq_f32(f32x4, vdupq_n_f32((float)zero_point)); s32x4 = vcvtaq_s32_f32(f32x4); s16x4_h = vqmovn_s32(s32x4); @@ -47,31 +49,37 @@ __float32_to_int8_neon_s8x8(float scale, float *input, int8_t *output) /* narrow to int8_t */ s8x8 = vqmovn_s16(s16x8); + s8x8 = vmax_s8(s8x8, vdup_n_s8(INT8_MIN + 1)); /* store 8 elements */ vst1_s8(output, s8x8); } static inline void -__float32_to_int8_neon_s8x1(float scale, float *input, int8_t *output) +__float32_to_int8_neon_s8x1(const float *input, int8_t *output, float scale, int8_t zero_point) { - int32_t s32; + float32x2_t f32x2; + int32x2_t s32x2; int16_t s16; /* scale and convert, round to nearest with ties away rounding mode */ - s32 = vcvtas_s32_f32(scale * (*input)); + f32x2 = vdiv_f32(vdup_n_f32(*input), vdup_n_f32(scale)); + f32x2 = vadd_f32(f32x2, vdup_n_f32((float)zero_point)); + s32x2 = vcvta_s32_f32(f32x2); + s32x2 = vmax_s32(s32x2, vdup_n_s32(INT8_MIN + 1)); /* saturate narrow */ - s16 = vqmovns_s32(s32); + s16 = vqmovns_s32(vget_lane_s32(s32x2, 0)); /* convert to int8_t */ *output = vqmovnh_s16(s16); } int -rte_ml_io_float32_to_int8(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_int8(const void *input, void *output, uint64_t nb_elements, float scale, + int8_t zero_point) { - float *input_buffer; + const float *input_buffer; int8_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -80,14 +88,14 @@ rte_ml_io_float32_to_int8(float scale, uint64_t nb_elements, void *input, void * if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (int8_t *)output; vlen = 2 * sizeof(float) / sizeof(int8_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __float32_to_int8_neon_s8x8(scale, input_buffer, output_buffer); + __float32_to_int8_neon_s8x8(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -95,7 +103,7 @@ rte_ml_io_float32_to_int8(float scale, uint64_t nb_elements, void *input, void * /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __float32_to_int8_neon_s8x1(scale, input_buffer, output_buffer); + __float32_to_int8_neon_s8x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -104,7 +112,7 @@ rte_ml_io_float32_to_int8(float scale, uint64_t nb_elements, void *input, void * } static inline void -__int8_to_float32_neon_f32x8(float scale, int8_t *input, float *output) +__int8_to_float32_neon_f32x8(const int8_t *input, float *output, float scale, int8_t zero_point) { float32x4_t f32x4; int16x8_t s16x8; @@ -122,6 +130,7 @@ __int8_to_float32_neon_f32x8(float scale, int8_t *input, float *output) s16x4 = vget_low_s16(s16x8); s32x4 = vmovl_s16(s16x4); f32x4 = vcvtq_f32_s32(s32x4); + f32x4 = vsubq_f32(f32x4, vdupq_n_f32((float)zero_point)); f32x4 = vmulq_n_f32(f32x4, scale); vst1q_f32(output, f32x4); @@ -129,20 +138,22 @@ __int8_to_float32_neon_f32x8(float scale, int8_t *input, float *output) s16x4 = vget_high_s16(s16x8); s32x4 = vmovl_s16(s16x4); f32x4 = vcvtq_f32_s32(s32x4); + f32x4 = vsubq_f32(f32x4, vdupq_n_f32((float)zero_point)); f32x4 = vmulq_n_f32(f32x4, scale); vst1q_f32(output + 4, f32x4); } static inline void -__int8_to_float32_neon_f32x1(float scale, int8_t *input, float *output) +__int8_to_float32_neon_f32x1(const int8_t *input, float *output, float scale, int8_t zero_point) { - *output = scale * vcvts_f32_s32((int32_t)*input); + *output = scale * (vcvts_f32_s32((int32_t)*input) - (float)zero_point); } int -rte_ml_io_int8_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_int8_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + int8_t zero_point) { - int8_t *input_buffer; + const int8_t *input_buffer; float *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -151,14 +162,14 @@ rte_ml_io_int8_to_float32(float scale, uint64_t nb_elements, void *input, void * if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (int8_t *)input; + input_buffer = (const int8_t *)input; output_buffer = (float *)output; vlen = 2 * sizeof(float) / sizeof(int8_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __int8_to_float32_neon_f32x8(scale, input_buffer, output_buffer); + __int8_to_float32_neon_f32x8(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -166,7 +177,7 @@ rte_ml_io_int8_to_float32(float scale, uint64_t nb_elements, void *input, void * /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __int8_to_float32_neon_f32x1(scale, input_buffer, output_buffer); + __int8_to_float32_neon_f32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -175,7 +186,7 @@ rte_ml_io_int8_to_float32(float scale, uint64_t nb_elements, void *input, void * } static inline void -__float32_to_uint8_neon_u8x8(float scale, float *input, uint8_t *output) +__float32_to_uint8_neon_u8x8(const float *input, uint8_t *output, float scale, uint8_t zero_point) { uint16x4_t u16x4_l; uint16x4_t u16x4_h; @@ -188,7 +199,8 @@ __float32_to_uint8_neon_u8x8(float scale, float *input, uint8_t *output) * use round to nearest with ties away rounding mode. */ f32x4 = vld1q_f32(input); - f32x4 = vmulq_n_f32(f32x4, scale); + f32x4 = vdivq_f32(f32x4, vdupq_n_f32(scale)); + f32x4 = vaddq_f32(f32x4, vdupq_n_f32((float)zero_point)); u32x4 = vcvtaq_u32_f32(f32x4); u16x4_l = vqmovn_u32(u32x4); @@ -196,7 +208,8 @@ __float32_to_uint8_neon_u8x8(float scale, float *input, uint8_t *output) * use round to nearest with ties away rounding mode. */ f32x4 = vld1q_f32(input + 4); - f32x4 = vmulq_n_f32(f32x4, scale); + f32x4 = vdivq_f32(f32x4, vdupq_n_f32(scale)); + f32x4 = vaddq_f32(f32x4, vdupq_n_f32((float)zero_point)); u32x4 = vcvtaq_u32_f32(f32x4); u16x4_h = vqmovn_u32(u32x4); @@ -211,25 +224,29 @@ __float32_to_uint8_neon_u8x8(float scale, float *input, uint8_t *output) } static inline void -__float32_to_uint8_neon_u8x1(float scale, float *input, uint8_t *output) +__float32_to_uint8_neon_u8x1(const float *input, uint8_t *output, float scale, uint8_t zero_point) { - uint32_t u32; + float32x2_t f32x2; + uint32x2_t u32x2; uint16_t u16; /* scale and convert, round to nearest with ties away rounding mode */ - u32 = vcvtas_u32_f32(scale * (*input)); + f32x2 = vdiv_f32(vdup_n_f32(*input), vdup_n_f32(scale)); + f32x2 = vadd_f32(f32x2, vdup_n_f32((float)zero_point)); + u32x2 = vcvta_u32_f32(f32x2); /* saturate narrow */ - u16 = vqmovns_u32(u32); + u16 = vqmovns_u32(vget_lane_u32(u32x2, 0)); /* convert to uint8_t */ *output = vqmovnh_u16(u16); } int -rte_ml_io_float32_to_uint8(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_uint8(const void *input, void *output, uint64_t nb_elements, float scale, + uint8_t zero_point) { - float *input_buffer; + const float *input_buffer; uint8_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -238,14 +255,14 @@ rte_ml_io_float32_to_uint8(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint8_t *)output; vlen = 2 * sizeof(float) / sizeof(uint8_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __float32_to_uint8_neon_u8x8(scale, input_buffer, output_buffer); + __float32_to_uint8_neon_u8x8(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -253,7 +270,7 @@ rte_ml_io_float32_to_uint8(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __float32_to_uint8_neon_u8x1(scale, input_buffer, output_buffer); + __float32_to_uint8_neon_u8x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -262,45 +279,48 @@ rte_ml_io_float32_to_uint8(float scale, uint64_t nb_elements, void *input, void } static inline void -__uint8_to_float32_neon_f32x8(float scale, uint8_t *input, float *output) +__uint8_to_float32_neon_f32x8(const uint8_t *input, float *output, float scale, uint8_t zero_point) { float32x4_t f32x4; uint16x8_t u16x8; - uint16x4_t u16x4; - uint32x4_t u32x4; + int16x8_t s16x8; + int16x4_t s16x4; + int32x4_t s32x4; uint8x8_t u8x8; /* load 8 x uint8_t elements */ u8x8 = vld1_u8(input); - - /* widen uint8_t to uint16_t */ u16x8 = vmovl_u8(u8x8); + s16x8 = vreinterpretq_s16_u16(u16x8); /* convert lower 4 elements: widen to uint32_t, convert to float, scale and store */ - u16x4 = vget_low_u16(u16x8); - u32x4 = vmovl_u16(u16x4); - f32x4 = vcvtq_f32_u32(u32x4); + s16x4 = vget_low_s16(s16x8); + s32x4 = vmovl_s16(s16x4); + f32x4 = vcvtq_f32_s32(s32x4); + f32x4 = vsubq_f32(f32x4, vdupq_n_f32((float)zero_point)); f32x4 = vmulq_n_f32(f32x4, scale); vst1q_f32(output, f32x4); /* convert higher 4 elements: widen to uint32_t, convert to float, scale and store */ - u16x4 = vget_high_u16(u16x8); - u32x4 = vmovl_u16(u16x4); - f32x4 = vcvtq_f32_u32(u32x4); + s16x4 = vget_high_s16(s16x8); + s32x4 = vmovl_s16(s16x4); + f32x4 = vcvtq_f32_s32(s32x4); + f32x4 = vsubq_f32(f32x4, vdupq_n_f32((float)zero_point)); f32x4 = vmulq_n_f32(f32x4, scale); vst1q_f32(output + 4, f32x4); } static inline void -__uint8_to_float32_neon_f32x1(float scale, uint8_t *input, float *output) +__uint8_to_float32_neon_f32x1(const uint8_t *input, float *output, float scale, uint8_t zero_point) { - *output = scale * vcvts_f32_u32((uint32_t)*input); + *output = scale * (vcvts_f32_u32((uint32_t)*input) - (float)zero_point); } int -rte_ml_io_uint8_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_uint8_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + uint8_t zero_point) { - uint8_t *input_buffer; + const uint8_t *input_buffer; float *output_buffer; uint64_t nb_iterations; uint64_t vlen; @@ -309,14 +329,14 @@ rte_ml_io_uint8_to_float32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint8_t *)input; + input_buffer = (const uint8_t *)input; output_buffer = (float *)output; vlen = 2 * sizeof(float) / sizeof(uint8_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __uint8_to_float32_neon_f32x8(scale, input_buffer, output_buffer); + __uint8_to_float32_neon_f32x8(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -324,7 +344,7 @@ rte_ml_io_uint8_to_float32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __uint8_to_float32_neon_f32x1(scale, input_buffer, output_buffer); + __uint8_to_float32_neon_f32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -333,7 +353,7 @@ rte_ml_io_uint8_to_float32(float scale, uint64_t nb_elements, void *input, void } static inline void -__float32_to_int16_neon_s16x4(float scale, float *input, int16_t *output) +__float32_to_int16_neon_s16x4(const float *input, int16_t *output, float scale, int16_t zero_point) { float32x4_t f32x4; int16x4_t s16x4; @@ -343,34 +363,43 @@ __float32_to_int16_neon_s16x4(float scale, float *input, int16_t *output) f32x4 = vld1q_f32(input); /* scale */ - f32x4 = vmulq_n_f32(f32x4, scale); + f32x4 = vdivq_f32(f32x4, vdupq_n_f32(scale)); + + /* add zero point */ + f32x4 = vaddq_f32(f32x4, vdupq_n_f32((float)zero_point)); /* convert to int32x4_t using round to nearest with ties away rounding mode */ s32x4 = vcvtaq_s32_f32(f32x4); /* saturate narrow to int16x4_t */ s16x4 = vqmovn_s32(s32x4); + s16x4 = vmax_s16(s16x4, vdup_n_s16(INT16_MIN + 1)); /* store 4 elements */ vst1_s16(output, s16x4); } static inline void -__float32_to_int16_neon_s16x1(float scale, float *input, int16_t *output) +__float32_to_int16_neon_s16x1(const float *input, int16_t *output, float scale, int16_t zero_point) { - int32_t s32; + float32x2_t f32x2; + int32x2_t s32x2; /* scale and convert, round to nearest with ties away rounding mode */ - s32 = vcvtas_s32_f32(scale * (*input)); + f32x2 = vdiv_f32(vdup_n_f32(*input), vdup_n_f32(scale)); + f32x2 = vadd_f32(f32x2, vdup_n_f32((float)zero_point)); + s32x2 = vcvta_s32_f32(f32x2); + s32x2 = vmax_s32(s32x2, vdup_n_s32(INT16_MIN + 1)); /* saturate narrow */ - *output = vqmovns_s32(s32); + *output = vqmovns_s32(vget_lane_s32(s32x2, 0)); } int -rte_ml_io_float32_to_int16(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_int16(const void *input, void *output, uint64_t nb_elements, float scale, + int16_t zero_point) { - float *input_buffer; + const float *input_buffer; int16_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -379,14 +408,14 @@ rte_ml_io_float32_to_int16(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (int16_t *)output; vlen = 2 * sizeof(float) / sizeof(int16_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __float32_to_int16_neon_s16x4(scale, input_buffer, output_buffer); + __float32_to_int16_neon_s16x4(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -394,7 +423,7 @@ rte_ml_io_float32_to_int16(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __float32_to_int16_neon_s16x1(scale, input_buffer, output_buffer); + __float32_to_int16_neon_s16x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -403,7 +432,7 @@ rte_ml_io_float32_to_int16(float scale, uint64_t nb_elements, void *input, void } static inline void -__int16_to_float32_neon_f32x4(float scale, int16_t *input, float *output) +__int16_to_float32_neon_f32x4(const int16_t *input, float *output, float scale, int16_t zero_point) { float32x4_t f32x4; int16x4_t s16x4; @@ -418,6 +447,9 @@ __int16_to_float32_neon_f32x4(float scale, int16_t *input, float *output) /* convert int32_t to float */ f32x4 = vcvtq_f32_s32(s32x4); + /* subtract zero point */ + f32x4 = vsubq_f32(f32x4, vdupq_n_f32((float)zero_point)); + /* scale */ f32x4 = vmulq_n_f32(f32x4, scale); @@ -426,15 +458,16 @@ __int16_to_float32_neon_f32x4(float scale, int16_t *input, float *output) } static inline void -__int16_to_float32_neon_f32x1(float scale, int16_t *input, float *output) +__int16_to_float32_neon_f32x1(const int16_t *input, float *output, float scale, int16_t zero_point) { - *output = scale * vcvts_f32_s32((int32_t)*input); + *output = scale * (vcvts_f32_s32((int32_t)*input) - (float)zero_point); } int -rte_ml_io_int16_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_int16_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + int16_t zero_point) { - int16_t *input_buffer; + const int16_t *input_buffer; float *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -443,14 +476,14 @@ rte_ml_io_int16_to_float32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (int16_t *)input; + input_buffer = (const int16_t *)input; output_buffer = (float *)output; vlen = 2 * sizeof(float) / sizeof(int16_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __int16_to_float32_neon_f32x4(scale, input_buffer, output_buffer); + __int16_to_float32_neon_f32x4(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -458,7 +491,7 @@ rte_ml_io_int16_to_float32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __int16_to_float32_neon_f32x1(scale, input_buffer, output_buffer); + __int16_to_float32_neon_f32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -467,7 +500,8 @@ rte_ml_io_int16_to_float32(float scale, uint64_t nb_elements, void *input, void } static inline void -__float32_to_uint16_neon_u16x4(float scale, float *input, uint16_t *output) +__float32_to_uint16_neon_u16x4(const float *input, uint16_t *output, float scale, + uint16_t zero_point) { float32x4_t f32x4; uint16x4_t u16x4; @@ -477,7 +511,10 @@ __float32_to_uint16_neon_u16x4(float scale, float *input, uint16_t *output) f32x4 = vld1q_f32(input); /* scale */ - f32x4 = vmulq_n_f32(f32x4, scale); + f32x4 = vdivq_f32(f32x4, vdupq_n_f32(scale)); + + /* add zero point */ + f32x4 = vaddq_f32(f32x4, vdupq_n_f32((float)zero_point)); /* convert using round to nearest with ties to away rounding mode */ u32x4 = vcvtaq_u32_f32(f32x4); @@ -490,21 +527,23 @@ __float32_to_uint16_neon_u16x4(float scale, float *input, uint16_t *output) } static inline void -__float32_to_uint16_neon_u16x1(float scale, float *input, uint16_t *output) +__float32_to_uint16_neon_u16x1(const float *input, uint16_t *output, float scale, + uint16_t zero_point) { uint32_t u32; /* scale and convert, round to nearest with ties away rounding mode */ - u32 = vcvtas_u32_f32(scale * (*input)); + u32 = vcvtas_u32_f32((*input) / scale + (float)zero_point); /* saturate narrow */ - *output = vqmovns_u32(u32); + *output = vqmovns_u32(u32) + zero_point; } int -rte_ml_io_float32_to_uint16(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_uint16(const void *input, void *output, uint64_t nb_elements, float scale, + uint16_t zero_point) { - float *input_buffer; + const float *input_buffer; uint16_t *output_buffer; uint64_t nb_iterations; uint64_t vlen; @@ -513,14 +552,14 @@ rte_ml_io_float32_to_uint16(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint16_t *)output; vlen = 2 * sizeof(float) / sizeof(uint16_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __float32_to_uint16_neon_u16x4(scale, input_buffer, output_buffer); + __float32_to_uint16_neon_u16x4(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -528,7 +567,7 @@ rte_ml_io_float32_to_uint16(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __float32_to_uint16_neon_u16x1(scale, input_buffer, output_buffer); + __float32_to_uint16_neon_u16x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -537,7 +576,8 @@ rte_ml_io_float32_to_uint16(float scale, uint64_t nb_elements, void *input, void } static inline void -__uint16_to_float32_neon_f32x4(float scale, uint16_t *input, float *output) +__uint16_to_float32_neon_f32x4(const uint16_t *input, float *output, float scale, + uint16_t zero_point) { float32x4_t f32x4; uint16x4_t u16x4; @@ -552,6 +592,9 @@ __uint16_to_float32_neon_f32x4(float scale, uint16_t *input, float *output) /* convert uint32_t to float */ f32x4 = vcvtq_f32_u32(u32x4); + /* subtract zero point */ + f32x4 = vsubq_f32(f32x4, vdupq_n_f32((float)zero_point)); + /* scale */ f32x4 = vmulq_n_f32(f32x4, scale); @@ -560,15 +603,17 @@ __uint16_to_float32_neon_f32x4(float scale, uint16_t *input, float *output) } static inline void -__uint16_to_float32_neon_f32x1(float scale, uint16_t *input, float *output) +__uint16_to_float32_neon_f32x1(const uint16_t *input, float *output, float scale, + uint16_t zero_point) { - *output = scale * vcvts_f32_u32((uint32_t)*input); + *output = scale * (vcvts_f32_u32((uint32_t)*input) - (float)zero_point); } int -rte_ml_io_uint16_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_uint16_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + uint16_t zero_point) { - uint16_t *input_buffer; + const uint16_t *input_buffer; float *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -577,14 +622,14 @@ rte_ml_io_uint16_to_float32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint16_t *)input; + input_buffer = (const uint16_t *)input; output_buffer = (float *)output; vlen = 2 * sizeof(float) / sizeof(uint16_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __uint16_to_float32_neon_f32x4(scale, input_buffer, output_buffer); + __uint16_to_float32_neon_f32x4(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -592,7 +637,7 @@ rte_ml_io_uint16_to_float32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __uint16_to_float32_neon_f32x1(scale, input_buffer, output_buffer); + __uint16_to_float32_neon_f32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -601,7 +646,7 @@ rte_ml_io_uint16_to_float32(float scale, uint64_t nb_elements, void *input, void } static inline void -__float32_to_int32_neon_s32x4(float scale, float *input, int32_t *output) +__float32_to_int32_neon_s32x4(const float *input, int32_t *output, float scale, int32_t zero_point) { float32x4_t f32x4; int32x4_t s32x4; @@ -610,26 +655,43 @@ __float32_to_int32_neon_s32x4(float scale, float *input, int32_t *output) f32x4 = vld1q_f32(input); /* scale */ - f32x4 = vmulq_n_f32(f32x4, scale); + f32x4 = vdivq_f32(f32x4, vdupq_n_f32(scale)); + + /* add zero point */ + f32x4 = vaddq_f32(f32x4, vdupq_n_f32((float)zero_point)); /* convert to int32x4_t using round to nearest with ties away rounding mode */ s32x4 = vcvtaq_s32_f32(f32x4); + /* add zero_point */ + s32x4 = vaddq_s32(s32x4, vdupq_n_s32(zero_point)); + s32x4 = vmaxq_s32(s32x4, vdupq_n_s32(INT32_MIN + 1)); + /* store 4 elements */ vst1q_s32(output, s32x4); } static inline void -__float32_to_int32_neon_s32x1(float scale, float *input, int32_t *output) +__float32_to_int32_neon_s32x1(const float *input, int32_t *output, float scale, int32_t zero_point) { + float32x2_t f32x2; + int32x2_t s32x2; + /* scale and convert, round to nearest with ties away rounding mode */ - *output = vcvtas_s32_f32(scale * (*input)); + f32x2 = vdiv_f32(vdup_n_f32(*input), vdup_n_f32(scale)); + f32x2 = vadd_f32(f32x2, vdup_n_f32((float)zero_point)); + s32x2 = vcvta_s32_f32(f32x2); + s32x2 = vmax_s32(s32x2, vdup_n_s32(INT16_MIN + 1)); + + /* saturate narrow */ + vst1_lane_s32(output, s32x2, 0); } int -rte_ml_io_float32_to_int32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_int32(const void *input, void *output, uint64_t nb_elements, float scale, + int32_t zero_point) { - float *input_buffer; + const float *input_buffer; int32_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -638,14 +700,14 @@ rte_ml_io_float32_to_int32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (int32_t *)output; vlen = 2 * sizeof(float) / sizeof(int32_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __float32_to_int32_neon_s32x4(scale, input_buffer, output_buffer); + __float32_to_int32_neon_s32x4(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -653,7 +715,7 @@ rte_ml_io_float32_to_int32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __float32_to_int32_neon_s32x1(scale, input_buffer, output_buffer); + __float32_to_int32_neon_s32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -662,7 +724,7 @@ rte_ml_io_float32_to_int32(float scale, uint64_t nb_elements, void *input, void } static inline void -__int32_to_float32_neon_f32x4(float scale, int32_t *input, float *output) +__int32_to_float32_neon_f32x4(const int32_t *input, float *output, float scale, int32_t zero_point) { float32x4_t f32x4; int32x4_t s32x4; @@ -673,6 +735,9 @@ __int32_to_float32_neon_f32x4(float scale, int32_t *input, float *output) /* convert int32_t to float */ f32x4 = vcvtq_f32_s32(s32x4); + /* subtract zero point */ + f32x4 = vsubq_f32(f32x4, vdupq_n_f32((float)zero_point)); + /* scale */ f32x4 = vmulq_n_f32(f32x4, scale); @@ -681,15 +746,16 @@ __int32_to_float32_neon_f32x4(float scale, int32_t *input, float *output) } static inline void -__int32_to_float32_neon_f32x1(float scale, int32_t *input, float *output) +__int32_to_float32_neon_f32x1(const int32_t *input, float *output, float scale, int32_t zero_point) { - *output = scale * vcvts_f32_s32(*input); + *output = scale * (vcvts_f32_s32(*input) - (float)zero_point); } int -rte_ml_io_int32_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_int32_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + int32_t zero_point) { - int32_t *input_buffer; + const int32_t *input_buffer; float *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -698,14 +764,14 @@ rte_ml_io_int32_to_float32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (int32_t *)input; + input_buffer = (const int32_t *)input; output_buffer = (float *)output; vlen = 2 * sizeof(float) / sizeof(int32_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __int32_to_float32_neon_f32x4(scale, input_buffer, output_buffer); + __int32_to_float32_neon_f32x4(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -713,7 +779,7 @@ rte_ml_io_int32_to_float32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __int32_to_float32_neon_f32x1(scale, input_buffer, output_buffer); + __int32_to_float32_neon_f32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -722,7 +788,8 @@ rte_ml_io_int32_to_float32(float scale, uint64_t nb_elements, void *input, void } static inline void -__float32_to_uint32_neon_u32x4(float scale, float *input, uint32_t *output) +__float32_to_uint32_neon_u32x4(const float *input, uint32_t *output, float scale, + uint32_t zero_point) { float32x4_t f32x4; uint32x4_t u32x4; @@ -731,7 +798,10 @@ __float32_to_uint32_neon_u32x4(float scale, float *input, uint32_t *output) f32x4 = vld1q_f32(input); /* scale */ - f32x4 = vmulq_n_f32(f32x4, scale); + f32x4 = vdivq_f32(f32x4, vdupq_n_f32(scale)); + + /* add zero point */ + f32x4 = vaddq_f32(f32x4, vdupq_n_f32((float)zero_point)); /* convert using round to nearest with ties to away rounding mode */ u32x4 = vcvtaq_u32_f32(f32x4); @@ -741,16 +811,18 @@ __float32_to_uint32_neon_u32x4(float scale, float *input, uint32_t *output) } static inline void -__float32_to_uint32_neon_u32x1(float scale, float *input, uint32_t *output) +__float32_to_uint32_neon_u32x1(const float *input, uint32_t *output, float scale, + uint32_t zero_point) { /* scale and convert, round to nearest with ties away rounding mode */ - *output = vcvtas_u32_f32(scale * (*input)); + *output = vcvtas_u32_f32((*input) / scale + (float)zero_point); } int -rte_ml_io_float32_to_uint32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_uint32(const void *input, void *output, uint64_t nb_elements, float scale, + uint32_t zero_point) { - float *input_buffer; + const float *input_buffer; uint32_t *output_buffer; uint64_t nb_iterations; uint64_t vlen; @@ -759,14 +831,14 @@ rte_ml_io_float32_to_uint32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint32_t *)output; vlen = 2 * sizeof(float) / sizeof(uint32_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __float32_to_uint32_neon_u32x4(scale, input_buffer, output_buffer); + __float32_to_uint32_neon_u32x4(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -774,7 +846,7 @@ rte_ml_io_float32_to_uint32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __float32_to_uint32_neon_u32x1(scale, input_buffer, output_buffer); + __float32_to_uint32_neon_u32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -783,7 +855,8 @@ rte_ml_io_float32_to_uint32(float scale, uint64_t nb_elements, void *input, void } static inline void -__uint32_to_float32_neon_f32x4(float scale, uint32_t *input, float *output) +__uint32_to_float32_neon_f32x4(const uint32_t *input, float *output, float scale, + uint32_t zero_point) { float32x4_t f32x4; uint32x4_t u32x4; @@ -794,6 +867,9 @@ __uint32_to_float32_neon_f32x4(float scale, uint32_t *input, float *output) /* convert uint32_t to float */ f32x4 = vcvtq_f32_u32(u32x4); + /* subtract zero point */ + f32x4 = vsubq_f32(f32x4, vdupq_n_f32((float)zero_point)); + /* scale */ f32x4 = vmulq_n_f32(f32x4, scale); @@ -802,15 +878,17 @@ __uint32_to_float32_neon_f32x4(float scale, uint32_t *input, float *output) } static inline void -__uint32_to_float32_neon_f32x1(float scale, uint32_t *input, float *output) +__uint32_to_float32_neon_f32x1(const uint32_t *input, float *output, float scale, + uint32_t zero_point) { - *output = scale * vcvts_f32_u32(*input); + *output = scale * (vcvts_f32_u32(*input) - (float)zero_point); } int -rte_ml_io_uint32_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_uint32_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + uint32_t zero_point) { - uint32_t *input_buffer; + const uint32_t *input_buffer; float *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -819,14 +897,14 @@ rte_ml_io_uint32_to_float32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint32_t *)input; + input_buffer = (const uint32_t *)input; output_buffer = (float *)output; vlen = 2 * sizeof(float) / sizeof(uint32_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __uint32_to_float32_neon_f32x4(scale, input_buffer, output_buffer); + __uint32_to_float32_neon_f32x4(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -834,7 +912,7 @@ rte_ml_io_uint32_to_float32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __uint32_to_float32_neon_f32x1(scale, input_buffer, output_buffer); + __uint32_to_float32_neon_f32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -843,55 +921,68 @@ rte_ml_io_uint32_to_float32(float scale, uint64_t nb_elements, void *input, void } static inline void -__float32_to_int64_neon_s64x2(float scale, float *input, int64_t *output) +__float32_to_int64_neon_s64x2(const float *input, int64_t *output, float scale, int64_t zero_point) { float32x2_t f32x2; float64x2_t f64x2; int64x2_t s64x2; + int64_t s64; /* load 2 x float elements */ f32x2 = vld1_f32(input); /* scale */ - f32x2 = vmul_n_f32(f32x2, scale); + f32x2 = vdiv_f32(f32x2, vdup_n_f32(scale)); + + /* add zero point */ + f32x2 = vadd_f32(f32x2, vdup_n_f32((float)zero_point)); /* convert to float64x2_t */ f64x2 = vcvt_f64_f32(f32x2); /* convert to int64x2_t */ s64x2 = vcvtaq_s64_f64(f64x2); + s64 = vgetq_lane_s64(s64x2, 0); + s64 = (s64 == INT64_MIN) ? INT64_MIN + 1 : s64; - /* store 2 elements */ - vst1q_s64(output, s64x2); + /* store lane 0 of int64x2_t */ + *output = s64; } static inline void -__float32_to_int64_neon_s64x1(float scale, float *input, int64_t *output) +__float32_to_int64_neon_s64x1(const float *input, int64_t *output, float scale, int64_t zero_point) { float32x2_t f32x2; float64x2_t f64x2; int64x2_t s64x2; + int64_t s64; /* load 1 x float element */ f32x2 = vdup_n_f32(*input); /* scale */ - f32x2 = vmul_n_f32(f32x2, scale); + f32x2 = vdiv_f32(f32x2, vdup_n_f32(scale)); + + /* add zero point */ + f32x2 = vadd_f32(f32x2, vdup_n_f32((float)zero_point)); /* convert to float64x2_t */ f64x2 = vcvt_f64_f32(f32x2); /* convert to int64x2_t */ s64x2 = vcvtaq_s64_f64(f64x2); + s64 = vgetq_lane_s64(s64x2, 0); + s64 = (s64 == INT64_MIN) ? INT64_MIN + 1 : s64; /* store lane 0 of int64x2_t */ - vst1q_lane_s64(output, s64x2, 0); + *output = s64; } int -rte_ml_io_float32_to_int64(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_int64(const void *input, void *output, uint64_t nb_elements, float scale, + int64_t zero_point) { - float *input_buffer; + const float *input_buffer; int64_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -900,14 +991,14 @@ rte_ml_io_float32_to_int64(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (int64_t *)output; vlen = 4 * sizeof(float) / sizeof(int64_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __float32_to_int64_neon_s64x2(scale, input_buffer, output_buffer); + __float32_to_int64_neon_s64x2(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -915,7 +1006,7 @@ rte_ml_io_float32_to_int64(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __float32_to_int64_neon_s64x1(scale, input_buffer, output_buffer); + __float32_to_int64_neon_s64x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -924,7 +1015,7 @@ rte_ml_io_float32_to_int64(float scale, uint64_t nb_elements, void *input, void } static inline void -__int64_to_float32_neon_f32x2(float scale, int64_t *input, float *output) +__int64_to_float32_neon_f32x2(const int64_t *input, float *output, float scale, int64_t zero_point) { int64x2_t s64x2; float64x2_t f64x2; @@ -939,6 +1030,9 @@ __int64_to_float32_neon_f32x2(float scale, int64_t *input, float *output) /* convert float64x2_t to float32x2_t */ f32x2 = vcvt_f32_f64(f64x2); + /* subtract zero_point */ + f32x2 = vsub_f32(f32x2, vdup_n_f32(zero_point)); + /* scale */ f32x2 = vmul_n_f32(f32x2, scale); @@ -947,7 +1041,7 @@ __int64_to_float32_neon_f32x2(float scale, int64_t *input, float *output) } static inline void -__int64_to_float32_neon_f32x1(float scale, int64_t *input, float *output) +__int64_to_float32_neon_f32x1(const int64_t *input, float *output, float scale, int64_t zero_point) { int64x2_t s64x2; float64x2_t f64x2; @@ -962,17 +1056,21 @@ __int64_to_float32_neon_f32x1(float scale, int64_t *input, float *output) /* convert float64x2_t to float32x2_t */ f32x2 = vcvt_f32_f64(f64x2); + /* subtract zero_point */ + f32x2 = vsub_f32(f32x2, vdup_n_f32(zero_point)); + /* scale */ f32x2 = vmul_n_f32(f32x2, scale); - /* store float32x2_t */ + /* store float32x2_t lane 0 */ vst1_lane_f32(output, f32x2, 0); } int -rte_ml_io_int64_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_int64_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + int64_t zero_point) { - int64_t *input_buffer; + const int64_t *input_buffer; float *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -981,14 +1079,14 @@ rte_ml_io_int64_to_float32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (int64_t *)input; + input_buffer = (const int64_t *)input; output_buffer = (float *)output; vlen = 4 * sizeof(float) / sizeof(int64_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __int64_to_float32_neon_f32x2(scale, input_buffer, output_buffer); + __int64_to_float32_neon_f32x2(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -996,7 +1094,7 @@ rte_ml_io_int64_to_float32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __int64_to_float32_neon_f32x1(scale, input_buffer, output_buffer); + __int64_to_float32_neon_f32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -1005,7 +1103,8 @@ rte_ml_io_int64_to_float32(float scale, uint64_t nb_elements, void *input, void } static inline void -__float32_to_uint64_neon_u64x2(float scale, float *input, uint64_t *output) +__float32_to_uint64_neon_u64x2(const float *input, uint64_t *output, float scale, + uint64_t zero_point) { float32x2_t f32x2; float64x2_t f64x2; @@ -1015,7 +1114,10 @@ __float32_to_uint64_neon_u64x2(float scale, float *input, uint64_t *output) f32x2 = vld1_f32(input); /* scale */ - f32x2 = vmul_n_f32(f32x2, scale); + f32x2 = vdiv_f32(f32x2, vdup_n_f32(scale)); + + /* add zero point */ + f32x2 = vadd_f32(f32x2, vdup_n_f32((float)zero_point)); /* convert to float64x2_t */ f64x2 = vcvt_f64_f32(f32x2); @@ -1028,7 +1130,8 @@ __float32_to_uint64_neon_u64x2(float scale, float *input, uint64_t *output) } static inline void -__float32_to_uint64_neon_u64x1(float scale, float *input, uint64_t *output) +__float32_to_uint64_neon_u64x1(const float *input, uint64_t *output, float scale, + uint64_t zero_point) { float32x2_t f32x2; float64x2_t f64x2; @@ -1038,7 +1141,10 @@ __float32_to_uint64_neon_u64x1(float scale, float *input, uint64_t *output) f32x2 = vld1_lane_f32(input, vdup_n_f32(0), 0); /* scale */ - f32x2 = vmul_n_f32(f32x2, scale); + f32x2 = vdiv_f32(f32x2, vdup_n_f32(scale)); + + /* add zero_point */ + f32x2 = vadd_f32(f32x2, vdup_n_f32((float)zero_point)); /* convert to float64x2_t */ f64x2 = vcvt_f64_f32(f32x2); @@ -1051,9 +1157,10 @@ __float32_to_uint64_neon_u64x1(float scale, float *input, uint64_t *output) } int -rte_ml_io_float32_to_uint64(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_uint64(const void *input, void *output, uint64_t nb_elements, float scale, + uint64_t zero_point) { - float *input_buffer; + const float *input_buffer; uint64_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -1062,14 +1169,14 @@ rte_ml_io_float32_to_uint64(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint64_t *)output; vlen = 4 * sizeof(float) / sizeof(uint64_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __float32_to_uint64_neon_u64x2(scale, input_buffer, output_buffer); + __float32_to_uint64_neon_u64x2(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -1077,7 +1184,7 @@ rte_ml_io_float32_to_uint64(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __float32_to_uint64_neon_u64x1(scale, input_buffer, output_buffer); + __float32_to_uint64_neon_u64x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -1086,7 +1193,8 @@ rte_ml_io_float32_to_uint64(float scale, uint64_t nb_elements, void *input, void } static inline void -__uint64_to_float32_neon_f32x2(float scale, uint64_t *input, float *output) +__uint64_to_float32_neon_f32x2(const uint64_t *input, float *output, float scale, + uint64_t zero_point) { uint64x2_t u64x2; float64x2_t f64x2; @@ -1101,6 +1209,9 @@ __uint64_to_float32_neon_f32x2(float scale, uint64_t *input, float *output) /* convert float64x2_t to float32x2_t */ f32x2 = vcvt_f32_f64(f64x2); + /* subtract zero_point */ + f32x2 = vsub_f32(f32x2, vdup_n_f32((float)zero_point)); + /* scale */ f32x2 = vmul_n_f32(f32x2, scale); @@ -1109,7 +1220,8 @@ __uint64_to_float32_neon_f32x2(float scale, uint64_t *input, float *output) } static inline void -__uint64_to_float32_neon_f32x1(float scale, uint64_t *input, float *output) +__uint64_to_float32_neon_f32x1(const uint64_t *input, float *output, float scale, + uint64_t zero_point) { uint64x2_t u64x2; float64x2_t f64x2; @@ -1124,17 +1236,21 @@ __uint64_to_float32_neon_f32x1(float scale, uint64_t *input, float *output) /* convert float64x2_t to float32x2_t */ f32x2 = vcvt_f32_f64(f64x2); + /* subtract zero_point */ + f32x2 = vsub_f32(f32x2, vdup_n_f32((float)zero_point)); + /* scale */ f32x2 = vmul_n_f32(f32x2, scale); - /* store float32x2_t */ + /* store float32x2_t lane 0 */ vst1_lane_f32(output, f32x2, 0); } int -rte_ml_io_uint64_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_uint64_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + uint64_t zero_point) { - uint64_t *input_buffer; + const uint64_t *input_buffer; float *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -1143,14 +1259,14 @@ rte_ml_io_uint64_to_float32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint64_t *)input; + input_buffer = (const uint64_t *)input; output_buffer = (float *)output; vlen = 4 * sizeof(float) / sizeof(uint64_t); nb_iterations = nb_elements / vlen; /* convert vlen elements in each iteration */ for (i = 0; i < nb_iterations; i++) { - __uint64_to_float32_neon_f32x2(scale, input_buffer, output_buffer); + __uint64_to_float32_neon_f32x2(input_buffer, output_buffer, scale, zero_point); input_buffer += vlen; output_buffer += vlen; } @@ -1158,7 +1274,7 @@ rte_ml_io_uint64_to_float32(float scale, uint64_t nb_elements, void *input, void /* convert leftover elements */ i = i * vlen; for (; i < nb_elements; i++) { - __uint64_to_float32_neon_f32x1(scale, input_buffer, output_buffer); + __uint64_to_float32_neon_f32x1(input_buffer, output_buffer, scale, zero_point); input_buffer++; output_buffer++; } @@ -1167,7 +1283,7 @@ rte_ml_io_uint64_to_float32(float scale, uint64_t nb_elements, void *input, void } static inline void -__float32_to_float16_neon_f16x4(float32_t *input, float16_t *output) +__float32_to_float16_neon_f16x4(const float32_t *input, float16_t *output) { float32x4_t f32x4; float16x4_t f16x4; @@ -1183,7 +1299,7 @@ __float32_to_float16_neon_f16x4(float32_t *input, float16_t *output) } static inline void -__float32_to_float16_neon_f16x1(float32_t *input, float16_t *output) +__float32_to_float16_neon_f16x1(const float32_t *input, float16_t *output) { float32x4_t f32x4; float16x4_t f16x4; @@ -1199,9 +1315,9 @@ __float32_to_float16_neon_f16x1(float32_t *input, float16_t *output) } int -rte_ml_io_float32_to_float16(uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_float16(const void *input, void *output, uint64_t nb_elements) { - float32_t *input_buffer; + const float32_t *input_buffer; float16_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -1210,7 +1326,7 @@ rte_ml_io_float32_to_float16(uint64_t nb_elements, void *input, void *output) if ((nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float32_t *)input; + input_buffer = (const float32_t *)input; output_buffer = (float16_t *)output; vlen = 2 * sizeof(float32_t) / sizeof(float16_t); nb_iterations = nb_elements / vlen; @@ -1234,7 +1350,7 @@ rte_ml_io_float32_to_float16(uint64_t nb_elements, void *input, void *output) } static inline void -__float16_to_float32_neon_f32x4(float16_t *input, float32_t *output) +__float16_to_float32_neon_f32x4(const float16_t *input, float32_t *output) { float16x4_t f16x4; float32x4_t f32x4; @@ -1250,7 +1366,7 @@ __float16_to_float32_neon_f32x4(float16_t *input, float32_t *output) } static inline void -__float16_to_float32_neon_f32x1(float16_t *input, float32_t *output) +__float16_to_float32_neon_f32x1(const float16_t *input, float32_t *output) { float16x4_t f16x4; float32x4_t f32x4; @@ -1266,9 +1382,9 @@ __float16_to_float32_neon_f32x1(float16_t *input, float32_t *output) } int -rte_ml_io_float16_to_float32(uint64_t nb_elements, void *input, void *output) +rte_ml_io_float16_to_float32(const void *input, void *output, uint64_t nb_elements) { - float16_t *input_buffer; + const float16_t *input_buffer; float32_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -1277,7 +1393,7 @@ rte_ml_io_float16_to_float32(uint64_t nb_elements, void *input, void *output) if ((nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float16_t *)input; + input_buffer = (const float16_t *)input; output_buffer = (float32_t *)output; vlen = 2 * sizeof(float32_t) / sizeof(float16_t); nb_iterations = nb_elements / vlen; diff --git a/lib/mldev/mldev_utils_neon_bfloat16.c b/lib/mldev/mldev_utils_neon_bfloat16.c index 8dec3fd8343..b8e68532669 100644 --- a/lib/mldev/mldev_utils_neon_bfloat16.c +++ b/lib/mldev/mldev_utils_neon_bfloat16.c @@ -18,7 +18,7 @@ #ifdef __ARM_FEATURE_BF16 static inline void -__float32_to_bfloat16_neon_f16x4(float32_t *input, bfloat16_t *output) +__float32_to_bfloat16_neon_f16x4(const float32_t *input, bfloat16_t *output) { float32x4_t f32x4; bfloat16x4_t bf16x4; @@ -34,7 +34,7 @@ __float32_to_bfloat16_neon_f16x4(float32_t *input, bfloat16_t *output) } static inline void -__float32_to_bfloat16_neon_f16x1(float32_t *input, bfloat16_t *output) +__float32_to_bfloat16_neon_f16x1(const float32_t *input, bfloat16_t *output) { float32x4_t f32x4; bfloat16x4_t bf16x4; @@ -50,9 +50,9 @@ __float32_to_bfloat16_neon_f16x1(float32_t *input, bfloat16_t *output) } int -rte_ml_io_float32_to_bfloat16(uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_bfloat16(const void *input, void *output, uint64_t nb_elements) { - float32_t *input_buffer; + const float32_t *input_buffer; bfloat16_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -61,7 +61,7 @@ rte_ml_io_float32_to_bfloat16(uint64_t nb_elements, void *input, void *output) if ((nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float32_t *)input; + input_buffer = (const float32_t *)input; output_buffer = (bfloat16_t *)output; vlen = 2 * sizeof(float32_t) / sizeof(bfloat16_t); nb_iterations = nb_elements / vlen; @@ -85,7 +85,7 @@ rte_ml_io_float32_to_bfloat16(uint64_t nb_elements, void *input, void *output) } static inline void -__bfloat16_to_float32_neon_f32x4(bfloat16_t *input, float32_t *output) +__bfloat16_to_float32_neon_f32x4(const bfloat16_t *input, float32_t *output) { bfloat16x4_t bf16x4; float32x4_t f32x4; @@ -101,7 +101,7 @@ __bfloat16_to_float32_neon_f32x4(bfloat16_t *input, float32_t *output) } static inline void -__bfloat16_to_float32_neon_f32x1(bfloat16_t *input, float32_t *output) +__bfloat16_to_float32_neon_f32x1(const bfloat16_t *input, float32_t *output) { bfloat16x4_t bf16x4; float32x4_t f32x4; @@ -117,9 +117,9 @@ __bfloat16_to_float32_neon_f32x1(bfloat16_t *input, float32_t *output) } int -rte_ml_io_bfloat16_to_float32(uint64_t nb_elements, void *input, void *output) +rte_ml_io_bfloat16_to_float32(const void *input, void *output, uint64_t nb_elements) { - bfloat16_t *input_buffer; + const bfloat16_t *input_buffer; float32_t *output_buffer; uint64_t nb_iterations; uint32_t vlen; @@ -128,7 +128,7 @@ rte_ml_io_bfloat16_to_float32(uint64_t nb_elements, void *input, void *output) if ((nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (bfloat16_t *)input; + input_buffer = (const bfloat16_t *)input; output_buffer = (float32_t *)output; vlen = 2 * sizeof(float32_t) / sizeof(bfloat16_t); nb_iterations = nb_elements / vlen; diff --git a/lib/mldev/mldev_utils_scalar.c b/lib/mldev/mldev_utils_scalar.c index 63a9900cc8c..e1fefdec3b4 100644 --- a/lib/mldev/mldev_utils_scalar.c +++ b/lib/mldev/mldev_utils_scalar.c @@ -10,9 +10,10 @@ */ int -rte_ml_io_float32_to_int8(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_int8(const void *input, void *output, uint64_t nb_elements, float scale, + int8_t zero_point) { - float *input_buffer; + const float *input_buffer; int8_t *output_buffer; uint64_t i; int i32; @@ -20,11 +21,11 @@ rte_ml_io_float32_to_int8(float scale, uint64_t nb_elements, void *input, void * if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (int8_t *)output; for (i = 0; i < nb_elements; i++) { - i32 = (int32_t)round((*input_buffer) * scale); + i32 = (int32_t)(round(*input_buffer / scale) + zero_point); if (i32 < INT8_MIN) i32 = INT8_MIN; @@ -42,20 +43,21 @@ rte_ml_io_float32_to_int8(float scale, uint64_t nb_elements, void *input, void * } int -rte_ml_io_int8_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_int8_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + int8_t zero_point) { - int8_t *input_buffer; + const int8_t *input_buffer; float *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (int8_t *)input; + input_buffer = (const int8_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = scale * (float)(*input_buffer); + *output_buffer = scale * (float)(*input_buffer - zero_point); input_buffer++; output_buffer++; @@ -65,9 +67,10 @@ rte_ml_io_int8_to_float32(float scale, uint64_t nb_elements, void *input, void * } int -rte_ml_io_float32_to_uint8(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_uint8(const void *input, void *output, uint64_t nb_elements, float scale, + uint8_t zero_point) { - float *input_buffer; + const float *input_buffer; uint8_t *output_buffer; int32_t i32; uint64_t i; @@ -75,11 +78,11 @@ rte_ml_io_float32_to_uint8(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint8_t *)output; for (i = 0; i < nb_elements; i++) { - i32 = (int32_t)round((*input_buffer) * scale); + i32 = (int32_t)(round(*input_buffer / scale) + zero_point); if (i32 < 0) i32 = 0; @@ -97,20 +100,21 @@ rte_ml_io_float32_to_uint8(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_uint8_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_uint8_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + uint8_t zero_point) { - uint8_t *input_buffer; + const uint8_t *input_buffer; float *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint8_t *)input; + input_buffer = (const uint8_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = scale * (float)(*input_buffer); + *output_buffer = scale * (float)(*input_buffer - zero_point); input_buffer++; output_buffer++; @@ -120,9 +124,10 @@ rte_ml_io_uint8_to_float32(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_float32_to_int16(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_int16(const void *input, void *output, uint64_t nb_elements, float scale, + int16_t zero_point) { - float *input_buffer; + const float *input_buffer; int16_t *output_buffer; int32_t i32; uint64_t i; @@ -130,11 +135,11 @@ rte_ml_io_float32_to_int16(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (int16_t *)output; for (i = 0; i < nb_elements; i++) { - i32 = (int32_t)round((*input_buffer) * scale); + i32 = (int32_t)(round(*input_buffer / scale) + zero_point); if (i32 < INT16_MIN) i32 = INT16_MIN; @@ -152,20 +157,21 @@ rte_ml_io_float32_to_int16(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_int16_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_int16_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + int16_t zero_point) { - int16_t *input_buffer; + const int16_t *input_buffer; float *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (int16_t *)input; + input_buffer = (const int16_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = scale * (float)(*input_buffer); + *output_buffer = scale * (float)(*input_buffer - zero_point); input_buffer++; output_buffer++; @@ -175,9 +181,10 @@ rte_ml_io_int16_to_float32(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_float32_to_uint16(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_uint16(const void *input, void *output, uint64_t nb_elements, float scale, + uint16_t zero_point) { - float *input_buffer; + const float *input_buffer; uint16_t *output_buffer; int32_t i32; uint64_t i; @@ -185,11 +192,11 @@ rte_ml_io_float32_to_uint16(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint16_t *)output; for (i = 0; i < nb_elements; i++) { - i32 = (int32_t)round((*input_buffer) * scale); + i32 = (int32_t)(round(*input_buffer / scale) + zero_point); if (i32 < 0) i32 = 0; @@ -207,20 +214,21 @@ rte_ml_io_float32_to_uint16(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_uint16_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_uint16_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + uint16_t zero_point) { - uint16_t *input_buffer; + const uint16_t *input_buffer; float *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint16_t *)input; + input_buffer = (const uint16_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = scale * (float)(*input_buffer); + *output_buffer = scale * (float)(*input_buffer - zero_point); input_buffer++; output_buffer++; @@ -230,20 +238,21 @@ rte_ml_io_uint16_to_float32(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_float32_to_int32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_int32(const void *input, void *output, uint64_t nb_elements, float scale, + int32_t zero_point) { - float *input_buffer; + const float *input_buffer; int32_t *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (int32_t *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = (int32_t)round((*input_buffer) * scale); + *output_buffer = (int32_t)(round(*input_buffer / scale) + zero_point); input_buffer++; output_buffer++; @@ -253,20 +262,21 @@ rte_ml_io_float32_to_int32(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_int32_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_int32_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + int32_t zero_point) { - int32_t *input_buffer; + const int32_t *input_buffer; float *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (int32_t *)input; + input_buffer = (const int32_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = scale * (float)(*input_buffer); + *output_buffer = scale * (float)(*input_buffer - zero_point); input_buffer++; output_buffer++; @@ -276,9 +286,10 @@ rte_ml_io_int32_to_float32(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_float32_to_uint32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_uint32(const void *input, void *output, uint64_t nb_elements, float scale, + uint32_t zero_point) { - float *input_buffer; + const float *input_buffer; uint32_t *output_buffer; int32_t i32; uint64_t i; @@ -286,11 +297,11 @@ rte_ml_io_float32_to_uint32(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint32_t *)output; for (i = 0; i < nb_elements; i++) { - i32 = (int32_t)round((*input_buffer) * scale); + i32 = (int32_t)(round(*input_buffer / scale) + zero_point); if (i32 < 0) i32 = 0; @@ -305,20 +316,21 @@ rte_ml_io_float32_to_uint32(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_uint32_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_uint32_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + uint32_t zero_point) { - uint32_t *input_buffer; + const uint32_t *input_buffer; float *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint32_t *)input; + input_buffer = (const uint32_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = scale * (float)(*input_buffer); + *output_buffer = scale * (float)(*input_buffer - zero_point); input_buffer++; output_buffer++; @@ -328,20 +340,21 @@ rte_ml_io_uint32_to_float32(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_float32_to_int64(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_int64(const void *input, void *output, uint64_t nb_elements, float scale, + int64_t zero_point) { - float *input_buffer; + const float *input_buffer; int64_t *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (int64_t *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = (int64_t)round((*input_buffer) * scale); + *output_buffer = (int64_t)(round(*input_buffer / scale) + zero_point); input_buffer++; output_buffer++; @@ -351,20 +364,21 @@ rte_ml_io_float32_to_int64(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_int64_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_int64_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + int64_t zero_point) { - int64_t *input_buffer; + const int64_t *input_buffer; float *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (int64_t *)input; + input_buffer = (const int64_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = scale * (float)(*input_buffer); + *output_buffer = scale * (float)(*input_buffer - zero_point); input_buffer++; output_buffer++; @@ -374,9 +388,10 @@ rte_ml_io_int64_to_float32(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_float32_to_uint64(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_uint64(const void *input, void *output, uint64_t nb_elements, float scale, + uint64_t zero_point) { - float *input_buffer; + const float *input_buffer; uint64_t *output_buffer; int64_t i64; uint64_t i; @@ -384,11 +399,11 @@ rte_ml_io_float32_to_uint64(float scale, uint64_t nb_elements, void *input, void if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint64_t *)output; for (i = 0; i < nb_elements; i++) { - i64 = (int64_t)round((*input_buffer) * scale); + i64 = (int64_t)(round(*input_buffer / scale) + zero_point); if (i64 < 0) i64 = 0; @@ -403,20 +418,21 @@ rte_ml_io_float32_to_uint64(float scale, uint64_t nb_elements, void *input, void } int -rte_ml_io_uint64_to_float32(float scale, uint64_t nb_elements, void *input, void *output) +rte_ml_io_uint64_to_float32(const void *input, void *output, uint64_t nb_elements, float scale, + uint64_t zero_point) { - uint64_t *input_buffer; + const uint64_t *input_buffer; float *output_buffer; uint64_t i; if ((scale == 0) || (nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint64_t *)input; + input_buffer = (const uint64_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { - *output_buffer = scale * (float)(*input_buffer); + *output_buffer = scale * (float)(*input_buffer - zero_point); input_buffer++; output_buffer++; @@ -548,16 +564,16 @@ __float32_to_float16_scalar_rtn(float x) } int -rte_ml_io_float32_to_float16(uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_float16(const void *input, void *output, uint64_t nb_elements) { - float *input_buffer; + const float *input_buffer; uint16_t *output_buffer; uint64_t i; if ((nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint16_t *)output; for (i = 0; i < nb_elements; i++) { @@ -632,16 +648,16 @@ __float16_to_float32_scalar_rtx(uint16_t f16) } int -rte_ml_io_float16_to_float32(uint64_t nb_elements, void *input, void *output) +rte_ml_io_float16_to_float32(const void *input, void *output, uint64_t nb_elements) { - uint16_t *input_buffer; + const uint16_t *input_buffer; float *output_buffer; uint64_t i; if ((nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint16_t *)input; + input_buffer = (const uint16_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { diff --git a/lib/mldev/mldev_utils_scalar_bfloat16.c b/lib/mldev/mldev_utils_scalar_bfloat16.c index 14374163131..3f93272518f 100644 --- a/lib/mldev/mldev_utils_scalar_bfloat16.c +++ b/lib/mldev/mldev_utils_scalar_bfloat16.c @@ -92,16 +92,16 @@ __float32_to_bfloat16_scalar_rtn(float x) } int -rte_ml_io_float32_to_bfloat16(uint64_t nb_elements, void *input, void *output) +rte_ml_io_float32_to_bfloat16(const void *input, void *output, uint64_t nb_elements) { - float *input_buffer; + const float *input_buffer; uint16_t *output_buffer; uint64_t i; if ((nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (float *)input; + input_buffer = (const float *)input; output_buffer = (uint16_t *)output; for (i = 0; i < nb_elements; i++) { @@ -174,16 +174,16 @@ __bfloat16_to_float32_scalar_rtx(uint16_t f16) } int -rte_ml_io_bfloat16_to_float32(uint64_t nb_elements, void *input, void *output) +rte_ml_io_bfloat16_to_float32(const void *input, void *output, uint64_t nb_elements) { - uint16_t *input_buffer; + const uint16_t *input_buffer; float *output_buffer; uint64_t i; if ((nb_elements == 0) || (input == NULL) || (output == NULL)) return -EINVAL; - input_buffer = (uint16_t *)input; + input_buffer = (const uint16_t *)input; output_buffer = (float *)output; for (i = 0; i < nb_elements; i++) { diff --git a/lib/mldev/rte_mldev.h b/lib/mldev/rte_mldev.h index 634af3d5e1a..8b595839056 100644 --- a/lib/mldev/rte_mldev.h +++ b/lib/mldev/rte_mldev.h @@ -1013,6 +1013,468 @@ rte_ml_model_params_update(int16_t dev_id, uint16_t model_id, void *buffer); /* IO operations */ +/** + * Convert a buffer containing numbers in single precision floating format (float32) to signed 8-bit + * integer format (INT8). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] i8 + * Output buffer to store INT8 numbers. Size of buffer is equal to (nb_elements * 1) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_int8(const void *fp32, void *i8, uint64_t nb_elements, float scale, + int8_t zero_point); + +/** + * Convert a buffer containing numbers in signed 8-bit integer format (INT8) to single precision + * floating format (float32). + * + * @param[in] i8 + * Input buffer containing INT8 numbers. Size of buffer is equal to (nb_elements * 1) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_int8_to_float32(const void *i8, void *fp32, uint64_t nb_elements, float scale, + int8_t zero_point); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to unsigned + * 8-bit integer format (UINT8). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] ui8 + * Output buffer to store UINT8 numbers. Size of buffer is equal to (nb_elements * 1) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_uint8(const void *fp32, void *ui8, uint64_t nb_elements, float scale, + uint8_t zero_point); + +/** + * Convert a buffer containing numbers in unsigned 8-bit integer format (UINT8) to single precision + * floating format (float32). + * + * @param[in] ui8 + * Input buffer containing UINT8 numbers. Size of buffer is equal to (nb_elements * 1) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_uint8_to_float32(const void *ui8, void *fp32, uint64_t nb_elements, float scale, + uint8_t zero_point); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to signed + * 16-bit integer format (INT16). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] i16 + * Output buffer to store INT16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_int16(const void *fp32, void *i16, uint64_t nb_elements, float scale, + int16_t zero_point); + +/** + * Convert a buffer containing numbers in signed 16-bit integer format (INT16) to single precision + * floating format (float32). + * + * @param[in] i16 + * Input buffer containing INT16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_int16_to_float32(const void *i16, void *fp32, uint64_t nb_elements, float scale, + int16_t zero_point); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to unsigned + * 16-bit integer format (UINT16). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] ui16 + * Output buffer to store UINT16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_uint16(const void *fp32, void *ui16, uint64_t nb_elements, float scale, + uint16_t zero_point); + +/** + * Convert a buffer containing numbers in unsigned 16-bit integer format (UINT16) to single + * precision floating format (float32). + * + * @param[in] ui16 + * Input buffer containing UINT16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_uint16_to_float32(const void *ui16, void *fp32, uint64_t nb_elements, float scale, + uint16_t zero_point); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to signed + * 32-bit integer format (INT32). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] i32 + * Output buffer to store INT32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_int32(const void *fp32, void *i32, uint64_t nb_elements, float scale, + int32_t zero_point); + +/** + * Convert a buffer containing numbers in signed 32-bit integer format (INT32) to single precision + * floating format (float32). + * + * @param[in] i32 + * Input buffer containing INT32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ + +__rte_experimental +int +rte_ml_io_int32_to_float32(const void *i32, void *fp32, uint64_t nb_elements, float scale, + int32_t zero_point); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to unsigned + * 32-bit integer format (UINT32). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] ui32 + * Output buffer to store UINT32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_uint32(const void *fp32, void *ui32, uint64_t nb_elements, float scale, + uint32_t zero_point); + +/** + * Convert a buffer containing numbers in unsigned 32-bit integer format (UINT32) to single + * precision floating format (float32). + * + * @param[in] ui32 + * Input buffer containing UINT32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_uint32_to_float32(const void *ui32, void *fp32, uint64_t nb_elements, float scale, + uint32_t zero_point); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to signed + * 64-bit integer format (INT64). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] i64 + * Output buffer to store INT64 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_int64(const void *fp32, void *i64, uint64_t nb_elements, float scale, + int64_t zero_point); + +/** + * Convert a buffer containing numbers in signed 64-bit integer format (INT64) to single precision + * floating format (float32). + * + * @param[in] i64 + * Input buffer containing INT64 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_int64_to_float32(const void *i64, void *fp32, uint64_t nb_elements, float scale, + int64_t zero_point); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to unsigned + * 64-bit integer format (UINT64). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] ui64 + * Output buffer to store UINT64 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_uint64(const void *fp32, void *ui64, uint64_t nb_elements, float scale, + uint64_t zero_point); + +/** + * Convert a buffer containing numbers in unsigned 64-bit integer format (UINT64) to single + *precision floating format (float32). + * + * @param[in] ui64 + * Input buffer containing UINT64 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * @param[in] scale + * Scale factor for conversion. + * @param[in] zero_point + * Zero point for conversion. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_uint64_to_float32(const void *ui64, void *fp32, uint64_t nb_elements, float scale, + uint64_t zero_point); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to half + * precision floating point format (FP16). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements *4) bytes. + * @param[out] fp16 + * Output buffer to store float16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_float16(const void *fp32, void *fp16, uint64_t nb_elements); + +/** + * Convert a buffer containing numbers in half precision floating format (FP16) to single precision + * floating point format (float32). + * + * @param[in] fp16 + * Input buffer containing float16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float16_to_float32(const void *fp16, void *fp32, uint64_t nb_elements); + +/** + * Convert a buffer containing numbers in single precision floating format (float32) to brain + * floating point format (bfloat16). + * + * @param[in] fp32 + * Input buffer containing float32 numbers. Size of buffer is equal to (nb_elements *4) bytes. + * @param[out] bf16 + * Output buffer to store bfloat16 numbers. Size of buffer is equal to (nb_elements * 2) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_float32_to_bfloat16(const void *fp32, void *bf16, uint64_t nb_elements); + +/** + * Convert a buffer containing numbers in brain floating point format (bfloat16) to single precision + * floating point format (float32). + * + * @param[in] bf16 + * Input buffer containing bfloat16 numbers. Size of buffer is equal to (nb_elements * 2) + * bytes. + * @param[out] fp32 + * Output buffer to store float32 numbers. Size of buffer is equal to (nb_elements * 4) bytes. + * @param[in] nb_elements + * Number of elements in the buffer. + * + * @return + * - 0, Success. + * - < 0, Error code on failure. + */ +__rte_experimental +int +rte_ml_io_bfloat16_to_float32(const void *bf16, void *fp32, uint64_t nb_elements); + /** * Quantize input data. * diff --git a/lib/mldev/version.map b/lib/mldev/version.map index 84bdd6c3004..6f0a70b4bd1 100644 --- a/lib/mldev/version.map +++ b/lib/mldev/version.map @@ -23,6 +23,26 @@ EXPERIMENTAL { rte_ml_dev_xstats_names_get; rte_ml_dev_xstats_reset; rte_ml_enqueue_burst; + rte_ml_io_float32_to_int8; + rte_ml_io_int8_to_float32; + rte_ml_io_float32_to_uint8; + rte_ml_io_uint8_to_float32; + rte_ml_io_float32_to_int16; + rte_ml_io_int16_to_float32; + rte_ml_io_float32_to_uint16; + rte_ml_io_uint16_to_float32; + rte_ml_io_float32_to_int32; + rte_ml_io_int32_to_float32; + rte_ml_io_float32_to_uint32; + rte_ml_io_uint32_to_float32; + rte_ml_io_float32_to_int64; + rte_ml_io_int64_to_float32; + rte_ml_io_float32_to_uint64; + rte_ml_io_uint64_to_float32; + rte_ml_io_float32_to_float16; + rte_ml_io_float16_to_float32; + rte_ml_io_float32_to_bfloat16; + rte_ml_io_bfloat16_to_float32; rte_ml_io_dequantize; rte_ml_io_quantize; rte_ml_model_info_get; @@ -50,24 +70,4 @@ INTERNAL { rte_ml_io_type_size_get; rte_ml_io_type_to_str; - rte_ml_io_float32_to_int8; - rte_ml_io_int8_to_float32; - rte_ml_io_float32_to_uint8; - rte_ml_io_uint8_to_float32; - rte_ml_io_float32_to_int16; - rte_ml_io_int16_to_float32; - rte_ml_io_float32_to_uint16; - rte_ml_io_uint16_to_float32; - rte_ml_io_float32_to_int32; - rte_ml_io_int32_to_float32; - rte_ml_io_float32_to_uint32; - rte_ml_io_uint32_to_float32; - rte_ml_io_float32_to_int64; - rte_ml_io_int64_to_float32; - rte_ml_io_float32_to_uint64; - rte_ml_io_uint64_to_float32; - rte_ml_io_float32_to_float16; - rte_ml_io_float16_to_float32; - rte_ml_io_float32_to_bfloat16; - rte_ml_io_bfloat16_to_float32; }; -- 2.45.1