zephyr/drivers/neural_net/intel_gna.h

204 lines
5.4 KiB
C

/*
* Copyright (c) 2018 Intel Corporation
*
* Author: Sathish Kuttan <sathish.k.kuttan@intel.com>
*
* SPDX-License-Identifier: Apache-2.0
*/
/** @file
* @brief Intel GNA device driver
*
* Device driver implementation for Intel's
* Gaussian Mixture Model and Neural Network Accelerator (GNA)
*/
#ifndef __INTEL_GNA__
#define __INTEL_GNA__
#ifdef __cplusplus
extern "C" {
#endif
/* number of requests that could be pending in driver */
#define GNA_REQUEST_QUEUE_LEN CONFIG_INTEL_GNA_MAX_PENDING_REQUESTS
#define GNA_MAX_NUM_MODELS CONFIG_INTEL_GNA_MAX_MODELS
/* values must match config values in Kconfig.intel_gna */
#define GNA_POWER_MODE_ALWAYS_ON 0
#define GNA_POWER_MODE_CLOCK_GATED 1
#define GNA_POWER_MODE_POWER_GATED 2
#define GNA_POWER_MODE_ALWAYS_OFF 3
#define INTEL_GNA_BASE_ADDR 0x0000E800
#define INTEL_GNA_IRQ_ID 0x00000506
#define INTEL_GNA_IRQ_PRIORITY 3
#define GNA_STS_INTR_PENDING BIT(31)
#define GNA_STS_SATURATION_OCCURRED BIT(17)
#define GNA_STS_BUFFER_FULL BIT(16)
#define GNA_STS_ERROR BIT(15)
#define GNA_STS_PARAM_OOR BIT(8)
#define GNA_STS_VIRT_ADDR_OOR BIT(7)
#define GNA_STS_STATS_VALID BIT(3)
#define GNA_STS_SUSP_PAUSE BIT(2)
#define GNA_STS_SUSP_BREAKP BIT(1)
#define GNA_STS_SCORE_COMPL BIT(0)
#define GNA_CTRL_INTR_DISABLE BIT(31)
#define GNA_CTRL_PM_IDLE_DISABLE BIT(18)
#define GNA_CTRL_PM_OVRIDE_CLK_ON BIT(17)
#define GNA_CTRL_PM_OVRIDE_PWR_ON BIT(16)
#define GNA_CTRL_STATS_ENABLE_STALL (1 << 12)
#define GNA_CTRL_STATS_MASK (BIT_MASK(4) << 12)
#define GNA_CTRL_ERR_INTR_ENABLE (1 << 10)
#define GNA_CTRL_COMPL_INTR_ENABLE (1 << 8)
#define GNA_CTRL_OPER_MODEL_XNN (1 << 5)
#define GNA_CTRL_ABORT_CLEAR (1 << 2)
#define GNA_CTRL_ACCEL_START (1 << 0)
#define GNA_CTRL_ACCEL_BUSY GNA_CTRL_ACCEL_START
#define GNA_CONFIG_DESC_PG_DIR_SIZE 64
#define GNA_LAYER_DESC_ALIGN (128)
#define GNA_ADDRESSABLE_MEM_SIZE L2_SRAM_SIZE
#define GNA_NUM_PG_TABLE_INDEX_BITS 10
#define GNA_NUM_PG_TABLE_ENTRIES BIT(GNA_NUM_PG_TABLE_INDEX_BITS)
#define GNA_PG_SIZE_IN_BITSHIFT 12
#define GNA_PG_SIZE_IN_BYTES BIT(GNA_PG_SIZE_IN_BITSHIFT)
#define GNA_SHIFT_RNDUP(value, shift) (((value) + BIT_MASK(shift)) >> (shift))
#define GNA_NUM_PAGES(bytes) \
GNA_SHIFT_RNDUP((bytes), GNA_PG_SIZE_IN_BITSHIFT)
#define GNA_PAGES_TO_BYTES(pages) ((pages) << GNA_PG_SIZE_IN_BITSHIFT)
#define GNA_MAX_NUM_PAGES GNA_NUM_PAGES(GNA_ADDRESSABLE_MEM_SIZE)
#define GNA_NUM_PG_TABLES_NEEDED \
GNA_SHIFT_RNDUP(GNA_MAX_NUM_PAGES, GNA_NUM_PG_TABLE_INDEX_BITS)
#if GNA_NUM_PG_TABLES_NEEDED > GNA_CONFIG_DESC_PG_DIR_SIZE
#error GNA_NUM_PG_TABLES_NEEDED exceeds GNA_CONFIG_DESC_PG_DIR_SIZE
#endif
#define GNA_GET_BITS(val, b_hi, b_lo) ((((uint32_t)(val)) << (31 - (b_hi))) >> \
(31 - (b_hi) + (b_lo)))
#define GNA_VA_PG_DIR(virt_addr) GNA_GET_BITS(virt_addr, 27, 22)
#define GNA_VA_PG_TABLE(virt_addr) GNA_GET_BITS(virt_addr, 21, 12)
#define GNA_PHYS_ADDR_TO_PAGE(addr) ((uint32_t)(addr) >> \
GNA_PG_SIZE_IN_BITSHIFT)
#define GNA_PG_DIR_ENTRY(phys_addr) GNA_PHYS_ADDR_TO_PAGE(phys_addr)
#define GNA_PG_BASE(addr) ((uint32_t)(addr) & \
~BIT_MASK(GNA_PG_SIZE_IN_BITSHIFT))
#define GNA_PG_OFFSET(addr) ((uint32_t)(addr) & \
BIT_MASK(GNA_PG_SIZE_IN_BITSHIFT))
#define GNA_PG_TABLE_ENTRY(phys_addr) GNA_PHYS_ADDR_TO_PAGE(phys_addr)
struct intel_gna_regs {
uint32_t gnasts;
uint32_t gnactrl;
uint32_t gnamctl;
uint32_t gnaptc;
uint32_t gnasc;
uint32_t gnaisi;
uint32_t gnais_low;
uint32_t gnais_high;
uint32_t gnabp_low;
uint32_t gnabp_high;
uint32_t reserved1[2];
uint32_t gnadesbase;
uint32_t gnaibuffs;
uint32_t reserved2[2];
uint32_t ovrcfgctl;
uint32_t reserved3[3];
uint32_t gnaversion;
};
struct intel_gna_config_desc {
uint32_t reserved1[64];
uint32_t labase; /* layer array base */
uint16_t lacnt; /* layer array count */
uint16_t reserved2;
uint32_t reserved3[62];
uint32_t vamaxaddr; /* virtual address max address */
uint32_t reserved4[3];
/* page directory entries */
uint32_t pagedir[GNA_CONFIG_DESC_PG_DIR_SIZE];
} __packed;
struct intel_gna_page_table {
uint32_t entry[GNA_NUM_PG_TABLE_ENTRIES];
} __aligned(GNA_PG_SIZE_IN_BYTES);
struct intel_gna_layer_desc {
uint32_t gna_words[8];
uint32_t inarrayptr;
uint32_t outarrayactptr;
uint32_t outarraysumptr;
uint32_t outfbarrayactptr;
uint32_t wtfltarrayptr;
uint32_t constarrayptr;
uint32_t actoutputslistptr;
uint32_t actfuncsectdefptr;
uint32_t reserved[16];
} __packed __aligned(GNA_LAYER_DESC_ALIGN);
struct intel_gna_config {
struct gna_config config;
};
struct intel_gna_model {
struct gna_model_info model;
void *input;
void *output;
void *vabase;
bool registered;
};
struct intel_gna_pending_req {
struct intel_gna_model *model;
void *output;
size_t output_len;
gna_callback callback;
};
struct intel_gna_pending_resp {
struct gna_inference_resp response;
gna_callback callback;
};
enum gna_state {
GNA_STATE_UNINITIALIZED,
GNA_STATE_INITIALIZED,
GNA_STATE_IDLE,
GNA_STATE_ACTIVE,
};
struct intel_gna_data {
/*
* gna_cb_work must be the first element in the structure
* since it will be typecast as intel_gna_data in the work handler
*/
struct k_work gna_work;
volatile struct intel_gna_regs *regs;
struct k_mem_slab model_slab;
struct intel_gna_model models[GNA_MAX_NUM_MODELS];
struct k_msgq request_queue;
struct intel_gna_pending_req requests[GNA_REQUEST_QUEUE_LEN];
struct k_msgq response_queue;
struct intel_gna_pending_resp responses[GNA_REQUEST_QUEUE_LEN];
enum gna_state state;
};
#ifdef __cplusplus
}
#endif
#endif /* __INTEL_GNA__ */