blob: d873832189242b54349c2f8d07338a9c1f2ca44c [file] [log] [blame]
#include <stdint.h>
#include <cpuinfo.h>
#include <log.h>
#include <arm/api.h>
#include <arm/midr.h>
void cpuinfo_arm_decode_cache(
enum cpuinfo_uarch uarch,
uint32_t cluster_cores,
uint32_t midr,
const struct cpuinfo_arm_chipset chipset[restrict static 1],
uint32_t cluster_id,
uint32_t arch_version,
struct cpuinfo_cache l1i[restrict static 1],
struct cpuinfo_cache l1d[restrict static 1],
struct cpuinfo_cache l2[restrict static 1])
{
switch (uarch) {
case cpuinfo_uarch_xscale:
switch (midr_get_part(midr) >> 8) {
case 2:
/*
* PXA 210/25X/26X
*
* See "Computer Organization and Design, Revised Printing: The Hardware/Software Interface"
* by David A. Patterson, John L. Hennessy
*/
*l1i = (struct cpuinfo_cache) {
.size = 16 * 1024,
.associativity = 32,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 16 * 1024,
.associativity = 4,
.line_size = 64
};
break;
case 4:
/* PXA 27X */
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 32,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 32,
.line_size = 32
};
break;
case 6:
/*
* PXA 3XX
*
* See http://download.intel.com/design/intelxscale/31628302.pdf
*/
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 32
};
*l2 = (struct cpuinfo_cache) {
.size = 256 * 1024,
.associativity = 8,
.line_size = 32
};
break;
}
break;
case cpuinfo_uarch_arm11:
*l1i = (struct cpuinfo_cache) {
.size = 16 * 1024,
.associativity = 4,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 16 * 1024,
.associativity = 4,
.line_size = 32
};
break;
case cpuinfo_uarch_cortex_a5:
/*
* Cortex-A5 Technical Reference Manual:
* 7.1.1. Memory system
* The Cortex-A5 processor has separate instruction and data caches.
* The caches have the following features:
* - Data cache is 4-way set-associative.
* - Instruction cache is 2-way set-associative.
* - The cache line length is eight words.
* - You can configure the instruction and data caches independently during implementation
* to sizes of 4KB, 8KB, 16KB, 32KB, or 64KB.
* 1.1.3. System design components
* PrimeCell Level 2 Cache Controller (PL310)
* The addition of an on-chip secondary cache, also referred to as a Level 2 or L2 cache, is a
* recognized method of improving the performance of ARM-based systems when significant memory traffic
* is generated by the processor. The PrimeCell Level 2 Cache Controller reduces the number of external
* memory accesses and has been optimized for use with the Cortex-A5 processor.
* 8.1.7. Exclusive L2 cache
* The Cortex-A5 processor can be connected to an L2 cache that supports an exclusive cache mode.
* This mode must be activated both in the Cortex-A5 processor and in the L2 cache controller.
*
* +--------------------+-----------+-----------+----------+-----------+
* | Processor model | L1D cache | L1I cache | L2 cache | Reference |
* +--------------------+-----------+-----------+----------+-----------+
* | Qualcomm MSM7225A | | | | |
* | Qualcomm MSM7625A | | | | |
* | Qualcomm MSM7227A | | | | |
* | Qualcomm MSM7627A | 32K | 32K | 256K | Wiki [1] |
* | Qualcomm MSM7225AB | | | | |
* | Qualcomm MSM7225AB | | | | |
* | Qualcomm QSD8250 | | | | |
* | Qualcomm QSD8650 | | | | |
* +--------------------+-----------+-----------+----------+-----------+
* | Spreadtrum SC6821 | 32K | 32K | ? | |
* | Spreadtrum SC6825 | 32K | 32K | 256K | Wiki [2] |
* | Spreadtrum SC8810 | ? | ? | ? | |
* | Spreadtrum SC8825 | 32K | 32K | ? | |
* +--------------------+-----------+-----------+----------+-----------+
*
* [1] https://en.wikipedia.org/wiki/List_of_Qualcomm_Snapdragon_systems-on-chip#Snapdragon_S1
* [2] https://en.wikipedia.org/wiki/Spreadtrum
*/
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 2,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 32
};
*l2 = (struct cpuinfo_cache) {
.size = 256 * 1024,
/*
* Follow NXP specification: "Eight-way set-associative 512 kB L2 cache with 32B line size"
* Reference: http://www.nxp.com/assets/documents/data/en/application-notes/AN4947.pdf
*/
.associativity = 8,
.line_size = 32
};
break;
case cpuinfo_uarch_cortex_a7:
/*
* Cortex-A7 MPCore Technical Reference Manual:
* 6.1. About the L1 memory system
* The L1 memory system consists of separate instruction and data caches. You can configure the
* instruction and data caches independently during implementation to sizes of 8KB, 16KB, 32KB, or 64KB.
*
* The L1 instruction memory system has the following features:
* - Instruction side cache line length of 32-bytes.
* - 2-way set-associative instruction cache.
*
* The L1 data memory system has the following features:
* - Data side cache line length of 64-bytes.
* - 4-way set-associative data cache.
*
* 7.1. About the L2 Memory system
* The L2 memory system consists of an:
* - Optional tightly-coupled L2 cache that includes:
* - Configurable L2 cache size of 128KB, 256KB, 512KB, and 1MB.
* - Fixed line length of 64 bytes
* - 8-way set-associative cache structure
*
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Allwinner A20 | 2 | 32K | 32K | 256K | [1] |
* | Allwinner A23 | 2 | 32K | 32K | 256K | [2] |
* | Allwinner A31 | 4 | 32K | 32K | 1M | [3] |
* | Allwinner A31s | 4 | 32K | 32K | 1M | [4] |
* | Allwinner A33 | 4 | 32K | 32K | 512K | [5] |
* | Allwinner A80 Octa | 4(+4) | 32K | 32K | 512K(+2M) | [6] |
* | Allwinner A81T | 8 | 32K | 32K | 1M | [7] |
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Broadcom BCM2836 | 4 | 32K | 32K | 512K | [8] |
* +--------------------+-------+-----------+-----------+-----------+-----------+
*
* [1] https://linux-sunxi.org/A20
* [2] https://linux-sunxi.org/A23
* [3] http://dl.linux-sunxi.org/A31/A3x_release_document/A31/IC/A31%20datasheet%20V1.3%2020131106.pdf
* [4] https://github.com/allwinner-zh/documents/blob/master/A31s/A31s_Datasheet_v1.5_20150510.pdf
* [5] http://dl.linux-sunxi.org/A33/A33_Datasheet_release1.0.pdf
* [6] https://linux-sunxi.org/images/1/10/A80_Datasheet_Revision_1.0_0404.pdf
* [7] http://dl.linux-sunxi.org/A83T/A83T_datasheet_Revision_1.1.pdf
* [8] https://www.raspberrypi.org/forums/viewtopic.php?t=98428
*/
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 2,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = 128 * 1024 * cluster_cores,
.associativity = 8,
.line_size = 64
};
break;
case cpuinfo_uarch_cortex_a8:
/*
* Cortex-A8 Technical Reference Manual:
* 7.1. About the L1 memory system
* The L1 memory system consists of separate instruction and data caches in a Harvard arrangement.
* The L1 memory system provides the core with:
* - fixed line length of 64 bytes
* - support for 16KB or 32KB caches
* - 4-way set associative cache structure
* 8.1. About the L2 memory system
* The L2 memory system is tightly coupled to the L1 data cache and L1 instruction cache.
* The key features of the L2 memory system include:
* - configurable cache size of 0KB, 128KB, 256KB, 512KB, and 1MB
* - fixed line length of 64 bytes
* - 8-way set associative cache structure
*
* +----------------------+-----------+-----------+-----------+-----------+
* | Processor model | L1D cache | L1I cache | L2 cache | Reference |
* +----------------------+-----------+-----------+-----------+-----------+
* | Exynos 3 Single 3110 | 32K | 32K | 512K | [1] |
* +----------------------+-----------+-----------+-----------+-----------+
* | TI DM 3730 | 32K | 32K | 256K | [2] |
* +----------------------+-----------+-----------+-----------+-----------+
*
* [1] https://en.wikichip.org/w/images/0/04/Exynos_3110.pdf
* [2] https://www.ti.com/lit/ds/symlink/dm3725.pdf
*/
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.associativity = 8,
.line_size = 64
};
switch (chipset->vendor) {
case cpuinfo_arm_chipset_vendor_samsung:
l2->size = 512 * 1024;
break;
default:
l2->size = 256 * 1024;
break;
}
break;
case cpuinfo_uarch_cortex_a9:
/*
* ARM Cortex‑A9 Technical Reference Manual:
* 7.1.1 Memory system
* The Cortex‑A9 processor has separate instruction and data caches.
* The caches have the following features:
* - Both caches are 4-way set-associative.
* - The cache line length is eight words.
* - You can configure the instruction and data caches independently during implementation
* to sizes of 16KB, 32KB, or 64KB.
* 8.1.5 Exclusive L2 cache
* The Cortex‑A9 processor can be connected to an L2 cache that supports an exclusive cache mode.
* This mode must be activated both in the Cortex‑A9 processor and in the L2 cache controller.
*
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Exynos 4 Dual 4210 | 2 | 32K | 32K | 1M | [1] |
* | Exynos 4 Dual 4212 | 2 | 32K | 32K | 1M | [2] |
* | Exynos 4 Quad 4412 | 4 | 32K | 32K | 1M | [3] |
* | Exynos 4 Quad 4415 | 4 | 32K | 32K | 1M | |
* | TI OMAP 4430 | 2 | 32K | 32K | 1M | [4] |
* | TI OMAP 4460 | 2 | 32K | 32K | 1M | [5] |
* +--------------------+-------+-----------+-----------+-----------+-----------+
*
* [1] http://www.samsung.com/global/business/semiconductor/file/product/Exynos_4_Dual_45nm_User_Manaul_Public_REV1.00-0.pdf
* [2] http://www.samsung.com/global/business/semiconductor/file/product/Exynos_4_Dual_32nm_User_Manaul_Public_REV100-0.pdf
* [3] http://www.samsung.com/global/business/semiconductor/file/product/Exynos_4_Quad_User_Manaul_Public_REV1.00-0.pdf
* [4] https://www.hotchips.org/wp-content/uploads/hc_archives/hc21/2_mon/HC21.24.400.ClientProcessors-Epub/HC21.24.421.Witt-OMAP4430.pdf
* [5] http://www.anandtech.com/show/5310/samsung-galaxy-nexus-ice-cream-sandwich-review/9
*/
/* Use Exynos 4 specs */
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 32
};
*l2 = (struct cpuinfo_cache) {
.size = 1024 * 1024,
/* OMAP4460 in Pandaboard ES has 16-way set-associative L2 cache */
.associativity = 16,
.line_size = 32
};
break;
case cpuinfo_uarch_cortex_a15:
/*
* 6.1. About the L1 memory system
* The L1 memory system consists of separate instruction and data caches.
* The L1 instruction memory system has the following features:
* - 32KB 2-way set-associative instruction cache.
* - Fixed line length of 64 bytes.
* The L1 data memory system has the following features:
* - 32KB 2-way set-associative data cache.
* - Fixed line length of 64 bytes.
* 7.1. About the L2 memory system
* The features of the L2 memory system include:
* - Configurable L2 cache size of 512KB, 1MB, 2MB and 4MB.
* - Fixed line length of 64 bytes.
* - 16-way set-associative cache structure.
*
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Exynos 5 Dual 5250 | 2 | 32K | 32K | 1M | [1] |
* | Exynos 5 Hexa 5260 | 2(+4) | 32K | 32K | 1M(+512K) | [2] |
* | Exynos 5 Octa 5410 | 4(+4) | 32K | 32K | 2M(+512K) | [3] |
* | Exynos 5 Octa 5420 | 4(+4) | 32K | 32K | 2M(+512K) | [3] |
* | Exynos 5 Octa 5422 | 4(+4) | 32K | 32K | 2M(+512K) | [3] |
* | Exynos 5 Octa 5430 | 4(+4) | 32K | 32K | 2M(+512K) | [3] |
* | Exynos 5 Octa 5800 | 4(+4) | 32K | 32K | 2M(+512K) | [3] |
* +--------------------+-------+-----------+-----------+-----------+-----------+
*
* [1] http://www.arndaleboard.org/wiki/downloads/supports/Exynos_5_Dual_User_Manaul_Public_REV1.00.pdf
* [2] http://www.yicsystem.com/wp-content/uploads/2014/08/Espresso5260P-Guide-Book.pdf
* [3] http://www.anandtech.com/show/6768/samsung-details-exynos-5-octa-architecture-power-at-isscc-13
*/
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 2,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 2,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = cluster_cores * 512 * 1024,
.associativity = 16,
.line_size = 64
};
break;
case cpuinfo_uarch_cortex_a17:
/*
* ARM Cortex-A17 MPCore Processor Technical Reference Manual:
* 6.1. About the L1 memory system
* The L1 memory system consists of separate instruction and data caches.
* The size of the instruction cache is implemented as either 32KB or 64KB.
* The size of the data cache is 32KB.
*
* The L1 instruction cache has the following features:
* - Instruction side cache line length of 64-bytes.
* - 4-way set-associative instruction cache.
*
* The L1 data cache has the following features:
* - Data side cache line length of 64-bytes.
* - 4-way set-associative data cache.
*
* 7.1. About the L2 Memory system
* An integrated L2 cache:
* - The cache size is implemented as either 256KB, 512KB, 1MB, 2MB, 4MB or 8MB.
* - A fixed line length of 64 bytes.
* - 16-way set-associative cache structure.
*
* +------------------+-------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +------------------+-------+-----------+-----------+-----------+-----------+
* | MediaTek MT6595 | 4(+4) | 32K | 32K | 2M(+512K) | [1] |
* +------------------+-------+-----------+-----------+-----------+-----------+
*
* [1] https://blog.osakana.net/archives/5268
*/
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = cluster_cores * 512 * 1024,
.associativity = 16,
.line_size = 64
};
break;
case cpuinfo_uarch_cortex_a35:
/*
* ARM Cortex‑A35 Processor Technical Reference Manual:
* 6.1. About the L1 memory system
* The L1 memory system includes several power-saving and performance-enhancing features.
* These include separate instruction and data caches, which can be configured
* independently during implementation to sizes of 8KB, 16KB, 32KB, or 64KB.
*
* L1 instruction-side memory system
* A dedicated instruction cache that:
* - is virtually indexed and physically tagged.
* - is 2-way set associative.
* - is configurable to be 8KB, 16KB, 32KB, or 64KB.
* - uses a cache line length of 64 bytes.
*
* L1 data-side memory system
* A dedicated data cache that:
* - is physically indexed and physically tagged.
* - is 4-way set associative.
* - is configurable to be 8KB, 16KB, 32KB, or 64KB.
* - uses a cache line length of 64 bytes.
*
* 7.1. About the L2 memory system
* The L2 cache is 8-way set associative.
* Further features of the L2 cache are:
* - Configurable size of 128KB, 256KB, 512KB, and 1MB.
* - Fixed line length of 64 bytes.
* - Physically indexed and tagged.
*
* +-----------------+---------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +-----------------+---------+-----------+-----------+-----------+-----------+
* | MediaTek MT6599 | 4(+4+2) | ? | ? | ? | |
* +-----------------+---------+-----------+-----------+-----------+-----------+
*/
*l1i = (struct cpuinfo_cache) {
.size = 16 * 1024, /* assumption based on low-end Cortex-A53 */
.associativity = 2,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 16 * 1024, /* assumption based on low-end Cortex-A53 */
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = 256 * 1024, /* assumption based on low-end Cortex-A53 */
.associativity = 8,
.line_size = 64
};
break;
case cpuinfo_uarch_cortex_a53:
/*
* ARM Cortex-A53 MPCore Processor Technical Reference Manual:
* 6.1. About the L1 memory system
* The L1 memory system consists of separate instruction and data caches. The implementer configures the
* instruction and data caches independently during implementation, to sizes of 8KB, 16KB, 32KB, or 64KB.
*
* The L1 Instruction memory system has the following key features:
* - Instruction side cache line length of 64 bytes.
* - 2-way set associative L1 Instruction cache.
*
* The L1 Data memory system has the following features:
* - Data side cache line length of 64 bytes.
* - 4-way set associative L1 Data cache.
*
* 7.1. About the L2 memory system
* The L2 memory system consists of an:
* - Optional tightly-coupled L2 cache that includes:
* - Configurable L2 cache size of 128KB, 256KB, 512KB, 1MB and 2MB.
* - Fixed line length of 64 bytes.
* - 16-way set-associative cache structure.
*
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Broadcom BCM2837 | 4 | 16K | 16K | 512K | [1] |
* | Exynos 7420 | 4(+4) | 32K | 32K | 256K | [2, 3] |
* | Exynos 8890 | 4(+4) | 32K | 32K | 256K | [4] |
* | Snapdragon 410 | 4 | 32K | 32K | 512K | [3] |
* | Snapdragon 835 | 4(+4) | 32K+64K | 32K+64K | 1M(+2M) | sysfs |
* | Kirin 620 | 4+4 | 32K | 32K | 512K | [5] |
* +--------------------+-------+-----------+-----------+-----------+-----------+
*
* [1] https://www.raspberrypi.org/forums/viewtopic.php?f=91&t=145766
* [2] http://www.anandtech.com/show/9330/exynos-7420-deep-dive/2
* [3] https://www.usenix.org/system/files/conference/usenixsecurity16/sec16_paper_lipp.pdf
* [4] http://www.boardset.com/products/products_v8890.php
* [5] http://mirror.lemaker.org/Hi6220V100_Multi-Mode_Application_Processor_Function_Description.pdf
*/
if (midr_is_qualcomm_cortex_a53_silver(midr)) {
/* Qualcomm-modified Cortex-A53 in Snapdragon 630/660/835 */
uint32_t l2_size = 512 * 1024;
if (chipset->series == cpuinfo_arm_chipset_series_qualcomm_msm && chipset->model == 8998) {
/* Snapdragon 835 (MSM8998): 1 MB L2 (little cores only) */
l2_size = 1024 * 1024;
} else if (chipset->series == cpuinfo_arm_chipset_series_qualcomm_snapdragon && chipset->model == 630 && cluster_id == 0) {
/* Snapdragon 630 (MSM8998): 1 MB L2 for the big cores */
l2_size = 1024 * 1024;
}
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 2,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = l2_size,
.associativity = 16,
.line_size = 64
};
} else {
/* Standard Cortex-A53 */
/* Use conservative values by default */
uint32_t l1_size = 16 * 1024;
uint32_t l2_size = 256 * 1024;
switch (chipset->series) {
case cpuinfo_arm_chipset_series_qualcomm_msm:
l1_size = 32 * 1024;
l2_size = 512 * 1024;
switch (chipset->model) {
case 8937: /* Snapdragon 430 */
case 8940: /* Snapdragon 435 */
case 8953: /* Snapdragon 625 or 626 (8953PRO) */
if (cluster_id == 0) {
/* 1M L2 for big cluster */
l2_size = 1024 * 1024;
}
break;
case 8952: /* Snapdragon 617 */
if (cluster_id != 0) {
/* 256K L2 for LITTLE cluster */
l2_size = 256 * 1024;
}
break;
default:
/* Silence compiler warning about unhandled enum values */
break;
}
break;
case cpuinfo_arm_chipset_series_qualcomm_apq:
l1_size = 32 * 1024;
l2_size = 512 * 1024;
break;
case cpuinfo_arm_chipset_series_qualcomm_snapdragon:
l1_size = 32 * 1024;
l2_size = 512 * 1024;
if (chipset->model == 450 && cluster_id == 0) {
/* Snapdragon 450: 1M L2 for big cluster */
l2_size = 1024 * 1024;
}
break;
case cpuinfo_arm_chipset_series_hisilicon_hi:
case cpuinfo_arm_chipset_series_hisilicon_kirin:
l1_size = 32 * 1024;
l2_size = 512 * 1024;
break;
case cpuinfo_arm_chipset_series_samsung_exynos:
l1_size = 32 * 1024;
break;
default:
/* Silence compiler warning about unhandled enum values */
break;
}
*l1i = (struct cpuinfo_cache) {
.size = l1_size,
.associativity = 2,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = l1_size,
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = l2_size,
.associativity = 16,
.line_size = 64
};
}
break;
case cpuinfo_uarch_cortex_a57:
/*
* ARM Cortex-A57 MPCore Processor Technical Reference Manual:
* 6.1. About the L1 memory system
* The L1 memory system consists of separate instruction and data caches.
*
* The L1 instruction memory system has the following features:
* - 48KB 3-way set-associative instruction cache.
* - Fixed line length of 64 bytes.
*
* The L1 data memory system has the following features:
* - 32KB 2-way set-associative data cache.
* - Fixed line length of 64 bytes.
*
* 7.1 About the L2 memory system
* The features of the L2 memory system include:
* - Configurable L2 cache size of 512KB, 1MB, and 2MB.
* - Fixed line length of 64 bytes.
* - 16-way set-associative cache structure.
* - Inclusion property with L1 data caches.
*
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Snapdragon 810 | 4(+4) | 32K | 48K | 2M | [1] |
* | Exynos 7420 | 4(+4) | 32K | 48K | 2M | [2] |
* | Jetson TX1 | 4 | 32K | 48K | 2M | [3] |
* +--------------------+-------+-----------+-----------+-----------+-----------+
*
* [1] http://www.anandtech.com/show/9837/snapdragon-820-preview
* [2] http://www.anandtech.com/show/9330/exynos-7420-deep-dive/2
* [3] https://devblogs.nvidia.com/parallelforall/jetson-tx2-delivers-twice-intelligence-edge/
*/
*l1i = (struct cpuinfo_cache) {
.size = 48 * 1024,
.associativity = 3,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 2,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = cluster_cores * 512 * 1024,
.associativity = 16,
.line_size = 64,
.flags = CPUINFO_CACHE_INCLUSIVE
};
break;
case cpuinfo_uarch_cortex_a72:
/*
* ARM Cortex-A72 MPCore Processor Technical Reference Manual
* 6.1. About the L1 memory system
* The L1 memory system consists of separate instruction and data caches.
*
* The L1 instruction memory system has the following features:
* - 48KB 3-way set-associative instruction cache.
* - Fixed line length of 64 bytes.
*
* The L1 data memory system has the following features:
* - 32KB 2-way set-associative data cache.
* - Fixed cache line length of 64 bytes.
*
* 7.1 About the L2 memory system
* The features of the L2 memory system include:
* - Configurable L2 cache size of 512KB, 1MB, 2MB and 4MB.
* - Fixed line length of 64 bytes.
* - Banked pipeline structures.
* - Inclusion property with L1 data caches.
* - 16-way set-associative cache structure.
*
* +---------------------+---------+-----------+-----------+------------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +---------------------+---------+-----------+-----------+------------+-----------+
* | Snapdragon 650 | 2(+4) | 32K(+32K) | 48K(+32K) | 1M(+512K) | [1] |
* | Snapdragon 652 | 4(+4) | 32K(+32K) | 48K(+32K) | 1M(+512K) | [2] |
* | Snapdragon 653 | 4(+4) | 32K(+32K) | 48K(+32K) | 1M(+512K) | [3] |
* | HiSilicon Kirin 950 | 4(+4) | 32K+32K | 48K+32K | ? | |
* | HiSilicon Kirin 955 | 4(+4) | 32K+32K | 48K+32K | ? | |
* | MediaTek Helio X20 | 2(+4+4) | ? | ? | ? | |
* | MediaTek Helio X23 | 2(+4+4) | ? | ? | ? | |
* | MediaTek Helio X25 | 2(+4+4) | ? | ? | ? | |
* | MediaTek Helio X27 | 2(+4+4) | ? | ? | ? | |
* +---------------------+---------+-----------+-----------+------------+-----------+
*
* [1] http://pdadb.net/index.php?m=processor&id=578&c=qualcomm_snapdragon_618_msm8956__snapdragon_650
* [2] http://pdadb.net/index.php?m=processor&id=667&c=qualcomm_snapdragon_620_apq8076__snapdragon_652
* [3] http://pdadb.net/index.php?m=processor&id=692&c=qualcomm_snapdragon_653_msm8976sg__msm8976_pro
*/
*l1i = (struct cpuinfo_cache) {
.size = 48 * 1024,
.associativity = 3,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 2,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = 1024 * 1024,
.associativity = 16,
.line_size = 64,
.flags = CPUINFO_CACHE_INCLUSIVE
};
break;
case cpuinfo_uarch_cortex_a73:
{
/*
* ARM Cortex‑A73 MPCore Processor Technical Reference Manual
* 6.1. About the L1 memory system
* The L1 memory system consists of separate instruction and data caches.
* The size of the instruction cache is 64KB.
* The size of the data cache is configurable to either 32KB or 64KB.
*
* The L1 instruction memory system has the following key features:
* - Virtually Indexed, Physically Tagged (VIPT), four-way set-associative instruction cache.
* - Fixed cache line length of 64 bytes.
*
* The L1 data memory system has the following features:
* - ...the data cache behaves like an eight-way set associative PIPT cache (for 32KB configurations)
* and a 16-way set associative PIPT cache (for 64KB configurations).
* - Fixed cache line length of 64 bytes.
*
* 7.1 About the L2 memory system
* The L2 memory system consists of:
* - A tightly-integrated L2 cache with:
* - A configurable size of 256KB, 512KB, 1MB, 2MB, 4MB, or 8MB.
* - A 16-way, set-associative structure.
* - A fixed line length of 64 bytes.
*
* The ARM Cortex A73 - Artemis Unveiled [1]
* "ARM still envisions that most vendors will choose to use configurations of 1 to
* 2MB in consumer products. The L2 cache is inclusive of the L1 cache. "
*
* +---------------------+---------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +---------------------+---------+-----------+-----------+-----------+-----------+
* | HiSilicon Kirin 960 | 4(+4) | 64K+32K | 64K+32K | ? | [2] |
* | MediaTek Helio X30 | 2(+4+4) | ? | 64K+ ? | ? | |
* | Snapdragon 835 | 4(+4) | 64K+32K | 64K+32K | 2M(+1M) | sysfs |
* | Snapdragon 660 | 4(+4) | 64K+32K | 64K+32K | 2M(+1M) | [3] |
* +---------------------+---------+-----------+-----------+-----------+-----------+
*
* [1] http://www.anandtech.com/show/10347/arm-cortex-a73-artemis-unveiled/2
* [2] http://www.anandtech.com/show/11088/hisilicon-kirin-960-performance-and-power/3
* [3] https://arstechnica.com/gadgets/2017/05/qualcomms-snapdragon-660-and-630-bring-more-high-end-features-to-midrange-chips/
*/
uint32_t l2_size = 1024 * 1024;
switch (midr) {
case UINT32_C(0x51AF8001): /* Kryo 280 Gold */
l2_size = 2 * 1024 * 1024;
break;
case UINT32_C(0x51AF8002): /* Kryo 260 Gold */
default:
break;
}
*l1i = (struct cpuinfo_cache) {
.size = 64 * 1024,
.associativity = 4,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 64 * 1024,
.associativity = 16,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = l2_size,
.associativity = 16,
.line_size = 64,
.flags = CPUINFO_CACHE_INCLUSIVE
};
break;
}
case cpuinfo_uarch_scorpion:
/*
* - "The CPU includes 32KB instruction and data caches as
* well as a complete memory-management unit (MMU) suitable
* for high-level operating systems. The CPU also has
* 256KB of SRAM that can be allocated in 64KB increments
* to level-two (L2) cache or tightly coupled memory (TCM)." [1]
* We interpret it as L2 cache being 4-way set-associative on single-core Scorpion.
* - L1 Data Cache = 32 KB. 32 B/line. [2]
* - L2 Cache = 256 KB. 128 B/line. [2]
* - 256 KB (single-core) or 512 KB (dual-core) L2 cache [3]
* - Single or dual-core configuration [3]
* - For L1 cache assume the same associativity as Krait
*
* [1] https://www.qualcomm.com/media/documents/files/linley-report-on-dual-core-snapdragon.pdf
* [2] http://www.7-cpu.com/cpu/Snapdragon.html
* [3] https://en.wikipedia.org/wiki/Scorpion_(CPU)
*/
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 32
};
*l2 = (struct cpuinfo_cache) {
.size = cluster_cores * 256 * 1024,
.associativity = 4,
.line_size = 128
};
break;
case cpuinfo_uarch_krait:
/*
* - L0 Data cache = 4 KB. 64 B/line, direct mapped [1]
* - L0 Instruction cache = 4 KB. [1]
* - L1 Data cache = 16 KB. 64 B/line, 4-way [1]
* - L1 Instruction cache = 16 KB, 4-way [1]
* - L2 Cache = 1 MB, 128 B/line, 8-way. Each core has fast access only to 512 KB of L2 cache. [1]
* - L2 = 1MB (dual core) or 2MB (quad core), 8-way set associative [2]
*
* [1] http://www.7-cpu.com/cpu/Krait.html
* [2] http://www.anandtech.com/show/4940/qualcomm-new-snapdragon-s4-msm8960-krait-architecture/2
*/
*l1i = (struct cpuinfo_cache) {
.size = 16 * 1024,
.associativity = 4,
.line_size = 64 /* assume same as L1D */
};
*l1d = (struct cpuinfo_cache) {
.size = 16 * 1024,
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = cluster_cores * 512 * 1024,
.associativity = 8,
.line_size = 128
};
break;
case cpuinfo_uarch_kryo:
/*
* +-----------------+-------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +-----------------+-------+-----------+-----------+-----------+-----------+
* | Snapdragon 820 | 2+2 | ? | ? | 1M+512K | [1] |
* | Snapdragon 821 | 2+2 | ? | ? | 1M+512K | [1] |
* +-----------------+-------+-----------+-----------+-----------+-----------+
*
* [1] http://www.anandtech.com/show/9837/snapdragon-820-preview/2
*/
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024 /* TODO: verify */,
.associativity = 4,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 24 * 1024 /* TODO: verify */,
.associativity = 3,
.line_size = 64
};
if (midr_is_kryo_silver(midr)) {
/* Kryo "Silver" */
*l2 = (struct cpuinfo_cache) {
.size = 512 * 1024,
.associativity = 8,
.line_size = 128
};
} else {
/* Kryo "Gold" */
*l2 = (struct cpuinfo_cache) {
.size = 1024 * 1024,
.associativity = 8,
.line_size = 128
};
}
break;
case cpuinfo_uarch_denver:
/*
* The Denver chip includes a 128KB, 4-way level 1 instruction cache, a 64KB, 4-way level 2 data cache,
* and a 2MB, 16-way level 2 cache, all of which can service both cores. [1]
*
* All the caches have 64-byte lines. [2]
*
* [1] http://www.pcworld.com/article/2463900/nvidia-reveals-pc-like-performance-for-denver-tegra-k1.html
* [2] http://linleygroup.com/newsletters/newsletter_detail.php?num=5205&year=2014
*/
*l1i = (struct cpuinfo_cache) {
.size = 128 * 1024,
.associativity = 4,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 64 * 1024,
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = 2 * 1024 * 1024,
.associativity = 16,
.line_size = 64
};
break;
case cpuinfo_uarch_mongoose:
/*
* - "Moving past branch prediction we can see some elements of how the cache is set up for the L1 I$,
* namely 64 KB split into four sets with 128-byte line sizes for 128 cache lines per set" [1]
* - "For loads and stores, a 32 KB, 8-way set associative cache with 64 byte line size is used" [1]
* - "The L2 cache here is 2MB shared across all cores split into 16 sets. This memory is also split
* into 4 banks and has a 22 cycle latency" [1]
*
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Processor model | Cores | L1D cache | L1I cache | L2 cache | Reference |
* +--------------------+-------+-----------+-----------+-----------+-----------+
* | Exynos 8 Octa 8890 | 4(+4) | 64K | 32K | 2M | [1] |
* | Exynos 8 Octa 8895 | 4(+4) | 64K | 32K | 2M | [2] |
* +--------------------+-------+-----------+-----------+-----------+-----------+
*
* [1] http://www.anandtech.com/show/10590/hot-chips-2016-exynos-m1-architecture-disclosed
* [2] https://www.extremetech.com/mobile/244949-samsungs-exynos-8895-features-custom-cpu-cores-first-10nm-chip-market
*/
*l1i = (struct cpuinfo_cache) {
.size = 64 * 1024,
.associativity = 4,
.line_size = 128
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 8,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = 2 * 1024 * 1024,
.associativity = 16,
.line_size = 64
};
break;
case cpuinfo_uarch_thunderx:
/*
* "78K-Icache and 32K-D cache per core, 16 MB shared L2 cache" [1]
*
* [1] https://www.cavium.com/pdfFiles/ThunderX_CP_PB_Rev1.pdf
*/
*l1i = (struct cpuinfo_cache) {
.size = 78 * 1024,
.associativity = 4 /* assumption */,
.line_size = 64 /* assumption */
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4 /* assumption */,
.line_size = 64 /* assumption */
};
*l2 = (struct cpuinfo_cache) {
.size = 16 * 1024 * 1024,
.associativity = 8 /* assumption */,
.line_size = 64 /* assumption */
};
break;
case cpuinfo_uarch_cortex_a12:
case cpuinfo_uarch_cortex_a32:
default:
cpuinfo_log_warning("target uarch not recognized; using generic cache parameters");
/* Follow OpenBLAS */
if (arch_version >= 8) {
*l1i = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 64
};
*l1d = (struct cpuinfo_cache) {
.size = 32 * 1024,
.associativity = 4,
.line_size = 64
};
*l2 = (struct cpuinfo_cache) {
.size = cluster_cores * 256 * 1024,
.associativity = 8,
.line_size = 64
};
} else {
*l1i = (struct cpuinfo_cache) {
.size = 16 * 1024,
.associativity = 4,
.line_size = 32
};
*l1d = (struct cpuinfo_cache) {
.size = 16 * 1024,
.associativity = 4,
.line_size = 32
};
if (arch_version >= 7) {
*l2 = (struct cpuinfo_cache) {
.size = cluster_cores * 128 * 1024,
.associativity = 8,
.line_size = 32
};
}
}
break;
}
l1i->sets = l1i->size / (l1i->associativity * l1i->line_size);
l1i->partitions = 1;
l1d->sets = l1d->size / (l1d->associativity * l1d->line_size);
l1d->partitions = 1;
if (l2->size != 0) {
l2->sets = l2->size / (l2->associativity * l2->line_size);
l2->partitions = 1;
}
}