xref: /linux/drivers/accel/amdxdna/npu4_regs.c (revision b8ea7da314c2efcb9c2f559ed65b7a36c869d68e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2023-2024, Advanced Micro Devices, Inc.
4  */
5 
6 #include <drm/amdxdna_accel.h>
7 #include <drm/drm_device.h>
8 #include <drm/gpu_scheduler.h>
9 #include <linux/amd-pmf-io.h>
10 #include <linux/bits.h>
11 #include <linux/sizes.h>
12 
13 #include "aie2_pci.h"
14 #include "amdxdna_mailbox.h"
15 #include "amdxdna_pci_drv.h"
16 
17 /* NPU Public Registers on MpNPUAxiXbar (refer to Diag npu_registers.h) */
18 #define MPNPU_PWAITMODE                0x301003C
19 #define MPNPU_PUB_SEC_INTR             0x3010060
20 #define MPNPU_PUB_PWRMGMT_INTR         0x3010064
21 #define MPNPU_PUB_SCRATCH0             0x301006C
22 #define MPNPU_PUB_SCRATCH1             0x3010070
23 #define MPNPU_PUB_SCRATCH2             0x3010074
24 #define MPNPU_PUB_SCRATCH3             0x3010078
25 #define MPNPU_PUB_SCRATCH4             0x301007C
26 #define MPNPU_PUB_SCRATCH5             0x3010080
27 #define MPNPU_PUB_SCRATCH6             0x3010084
28 #define MPNPU_PUB_SCRATCH7             0x3010088
29 #define MPNPU_PUB_SCRATCH8             0x301008C
30 #define MPNPU_PUB_SCRATCH9             0x3010090
31 #define MPNPU_PUB_SCRATCH10            0x3010094
32 #define MPNPU_PUB_SCRATCH11            0x3010098
33 #define MPNPU_PUB_SCRATCH12            0x301009C
34 #define MPNPU_PUB_SCRATCH13            0x30100A0
35 #define MPNPU_PUB_SCRATCH14            0x30100A4
36 #define MPNPU_PUB_SCRATCH15            0x30100A8
37 #define MP0_C2PMSG_73                  0x3810A24
38 #define MP0_C2PMSG_123                 0x3810AEC
39 
40 #define MP1_C2PMSG_0                   0x3B10900
41 #define MP1_C2PMSG_60                  0x3B109F0
42 #define MP1_C2PMSG_61                  0x3B109F4
43 
44 #define MPNPU_SRAM_X2I_MAILBOX_0       0x3600000
45 #define MPNPU_SRAM_X2I_MAILBOX_15      0x361E000
46 #define MPNPU_SRAM_X2I_MAILBOX_31      0x363E000
47 #define MPNPU_SRAM_I2X_MAILBOX_31      0x363F000
48 
49 #define MMNPU_APERTURE0_BASE           0x3000000
50 #define MMNPU_APERTURE1_BASE           0x3600000
51 #define MMNPU_APERTURE3_BASE           0x3810000
52 #define MMNPU_APERTURE4_BASE           0x3B10000
53 
54 /* PCIe BAR Index for NPU4 */
55 #define NPU4_REG_BAR_INDEX	0
56 #define NPU4_MBOX_BAR_INDEX	0
57 #define NPU4_PSP_BAR_INDEX	4
58 #define NPU4_SMU_BAR_INDEX	5
59 #define NPU4_SRAM_BAR_INDEX	2
60 /* Associated BARs and Apertures */
61 #define NPU4_REG_BAR_BASE	MMNPU_APERTURE0_BASE
62 #define NPU4_MBOX_BAR_BASE	MMNPU_APERTURE0_BASE
63 #define NPU4_PSP_BAR_BASE	MMNPU_APERTURE3_BASE
64 #define NPU4_SMU_BAR_BASE	MMNPU_APERTURE4_BASE
65 #define NPU4_SRAM_BAR_BASE	MMNPU_APERTURE1_BASE
66 
67 #define NPU4_DPM_TOPS(ndev, hclk) (4096 * (ndev)->total_col * (hclk) / 1000000)
68 
69 const struct rt_config npu4_default_rt_cfg[] = {
70 	{ 5, 1, AIE2_RT_CFG_INIT }, /* PDI APP LOAD MODE */
71 	{ 10, 1, AIE2_RT_CFG_INIT }, /* DEBUG BUF */
72 	{ 14, 0, AIE2_RT_CFG_INIT, BIT_U64(AIE2_PREEMPT) }, /* Frame boundary preemption */
73 	{ 1, 1, AIE2_RT_CFG_CLK_GATING }, /* Clock gating on */
74 	{ 2, 1, AIE2_RT_CFG_CLK_GATING }, /* Clock gating on */
75 	{ 3, 1, AIE2_RT_CFG_CLK_GATING }, /* Clock gating on */
76 	{ 4, 1, AIE2_RT_CFG_CLK_GATING }, /* Clock gating on */
77 	{ 13, 0, AIE2_RT_CFG_FORCE_PREEMPT },
78 	{ 14, 0, AIE2_RT_CFG_FRAME_BOUNDARY_PREEMPT },
79 	{ 0 },
80 };
81 
82 const struct dpm_clk_freq npu4_dpm_clk_table[] = {
83 	{396, 792},
84 	{600, 1056},
85 	{792, 1152},
86 	{975, 1267},
87 	{975, 1267},
88 	{1056, 1408},
89 	{1152, 1584},
90 	{1267, 1800},
91 	{ 0 }
92 };
93 
94 const struct amdxdna_fw_feature_tbl npu4_fw_feature_table[] = {
95 	{ .major = 6, .min_minor = 12 },
96 	{ .features = BIT_U64(AIE2_PREEMPT), .major = 6, .min_minor = 12 },
97 	{ .features = BIT_U64(AIE2_TEMPORAL_ONLY), .major = 6, .min_minor = 12 },
98 	{ .features = BIT_U64(AIE2_NPU_COMMAND), .major = 6, .min_minor = 15 },
99 	{ .features = BIT_U64(AIE2_UPDATE_PROPERTY), .major = 6, .min_minor = 15 },
100 	{ .features = BIT_U64(AIE2_APP_HEALTH), .major = 6, .min_minor = 18 },
101 	{ .features = BIT_U64(AIE2_ADD_HOST_BUFFER), .major = 6, .min_minor = 18 },
102 	{ .features = BIT_U64(AIE2_GET_DEV_REVISION), .major = 6, .min_minor = 24 },
103 	{ .features = AIE2_ALL_FEATURES, .major = 7 },
104 	{ 0 }
105 };
106 
107 static int npu4_set_dpm(struct amdxdna_dev_hdl *ndev, u32 dpm_level)
108 {
109 	int ret;
110 
111 	ret = aie_smu_set_dpm(ndev->aie.smu_hdl, dpm_level);
112 	if (ret)
113 		return ret;
114 
115 	ndev->npuclk_freq = ndev->priv->dpm_clk_tbl[dpm_level].npuclk;
116 	ndev->hclk_freq = ndev->priv->dpm_clk_tbl[dpm_level].hclk;
117 	ndev->max_tops = NPU4_DPM_TOPS(ndev, ndev->priv->dpm_clk_tbl[ndev->max_dpm_level].hclk);
118 	ndev->curr_tops = NPU4_DPM_TOPS(ndev, ndev->hclk_freq);
119 
120 	XDNA_DBG(ndev->aie.xdna, "MP-NPU clock %d, H clock %d\n",
121 		 ndev->npuclk_freq, ndev->hclk_freq);
122 
123 	return 0;
124 }
125 
126 static int npu4_update_counters(struct amdxdna_dev_hdl *ndev)
127 {
128 	struct amd_pmf_npu_metrics npu_metrics;
129 	int ret;
130 
131 	ret = AIE2_GET_PMF_NPU_METRICS(&npu_metrics);
132 	if (ret)
133 		return ret;
134 
135 	ndev->npuclk_freq = npu_metrics.mpnpuclk_freq;
136 	ndev->hclk_freq = npu_metrics.npuclk_freq;
137 	ndev->curr_tops = NPU4_DPM_TOPS(ndev, ndev->hclk_freq);
138 
139 	return 0;
140 }
141 
142 const struct aie2_hw_ops npu4_hw_ops = {
143 	.set_dpm = npu4_set_dpm,
144 	.update_counters = npu4_update_counters,
145 };
146 
147 const struct amdxdna_rev_vbnv npu4_rev_vbnv_tbl[] = {
148 	{ AIE2_DEV_REVISION_STXA, "NPU Strix" },
149 	{ AIE2_DEV_REVISION_STXB, "NPU Strix" },
150 	{ AIE2_DEV_REVISION_KRK1, "NPU Krackan 1" },
151 	{ AIE2_DEV_REVISION_KRK2, "NPU Krackan 2" },
152 	{ AIE2_DEV_REVISION_HALO, "NPU Strix Halo" },
153 	{ AIE2_DEV_REVISION_GPT1, "NPU Gorgon Point 1" },
154 	{ AIE2_DEV_REVISION_GPT2, "NPU Gorgon Point 2" },
155 	{ AIE2_DEV_REVISION_GPT3, "NPU Gorgon Point 3" },
156 	{ 0 }
157 };
158 
159 static const struct amdxdna_dev_priv npu4_dev_priv = {
160 	.fw_path        = "amdnpu/17f0_10/",
161 	.rt_config	= npu4_default_rt_cfg,
162 	.dpm_clk_tbl	= npu4_dpm_clk_table,
163 	.col_align	= COL_ALIGN_NATURE,
164 	.col_opc	= 4096,
165 	.mbox_dev_addr  = NPU4_MBOX_BAR_BASE,
166 	.mbox_size      = 0, /* Use BAR size */
167 	.sram_dev_addr  = NPU4_SRAM_BAR_BASE,
168 	.hwctx_limit    = 16,
169 	.sram_offs      = {
170 		DEFINE_BAR_OFFSET(MBOX_CHANN_OFF, NPU4_SRAM, MPNPU_SRAM_X2I_MAILBOX_0),
171 		DEFINE_BAR_OFFSET(FW_ALIVE_OFF,   NPU4_SRAM, MPNPU_SRAM_X2I_MAILBOX_15),
172 	},
173 	.psp_regs_off   = {
174 		DEFINE_BAR_OFFSET(PSP_CMD_REG,    NPU4_PSP, MP0_C2PMSG_123),
175 		DEFINE_BAR_OFFSET(PSP_ARG0_REG,   NPU4_REG, MPNPU_PUB_SCRATCH3),
176 		DEFINE_BAR_OFFSET(PSP_ARG1_REG,   NPU4_REG, MPNPU_PUB_SCRATCH4),
177 		DEFINE_BAR_OFFSET(PSP_ARG2_REG,   NPU4_REG, MPNPU_PUB_SCRATCH9),
178 		DEFINE_BAR_OFFSET(PSP_INTR_REG,   NPU4_PSP, MP0_C2PMSG_73),
179 		DEFINE_BAR_OFFSET(PSP_STATUS_REG, NPU4_PSP, MP0_C2PMSG_123),
180 		DEFINE_BAR_OFFSET(PSP_RESP_REG,   NPU4_REG, MPNPU_PUB_SCRATCH3),
181 		DEFINE_BAR_OFFSET(PSP_PWAITMODE_REG, NPU4_REG, MPNPU_PWAITMODE),
182 	},
183 	.smu_regs_off   = {
184 		DEFINE_BAR_OFFSET(SMU_CMD_REG,  NPU4_SMU, MP1_C2PMSG_0),
185 		DEFINE_BAR_OFFSET(SMU_ARG_REG,  NPU4_SMU, MP1_C2PMSG_60),
186 		DEFINE_BAR_OFFSET(SMU_INTR_REG, NPU4_SMU, MMNPU_APERTURE4_BASE),
187 		DEFINE_BAR_OFFSET(SMU_RESP_REG, NPU4_SMU, MP1_C2PMSG_61),
188 		DEFINE_BAR_OFFSET(SMU_OUT_REG,  NPU4_SMU, MP1_C2PMSG_60),
189 	},
190 	.hw_ops		= &npu4_hw_ops
191 };
192 
193 const struct amdxdna_dev_info dev_npu4_info = {
194 	.reg_bar           = NPU4_REG_BAR_INDEX,
195 	.mbox_bar          = NPU4_MBOX_BAR_INDEX,
196 	.sram_bar          = NPU4_SRAM_BAR_INDEX,
197 	.psp_bar           = NPU4_PSP_BAR_INDEX,
198 	.smu_bar           = NPU4_SMU_BAR_INDEX,
199 	.first_col         = 0,
200 	.dev_mem_buf_shift = 15, /* 32 KiB aligned */
201 	.dev_mem_base      = AIE2_DEVM_BASE,
202 	.dev_mem_size      = AIE2_DEVM_SIZE,
203 	.default_vbnv      = "RyzenAI-npu4",
204 	.dev_heap_max_size = AIE2_DEVM_MAX_SIZE,
205 	.device_type       = AMDXDNA_DEV_TYPE_KMQ,
206 	.rev_vbnv_tbl      = npu4_rev_vbnv_tbl,
207 	.dev_priv          = &npu4_dev_priv,
208 	.fw_feature_tbl    = npu4_fw_feature_table,
209 	.ops               = &aie2_ops, /* NPU4 can share NPU1's callback */
210 };
211