2 * Copyright 2016 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
23 #include <linux/firmware.h>
26 #include "amdgpu_gfx.h"
30 #include "vega10/soc15ip.h"
31 #include "vega10/GC/gc_9_0_offset.h"
32 #include "vega10/GC/gc_9_0_sh_mask.h"
33 #include "vega10/vega10_enum.h"
34 #include "vega10/HDP/hdp_4_0_offset.h"
36 #include "soc15_common.h"
37 #include "clearstate_gfx9.h"
38 #include "v9_structs.h"
40 #define GFX9_NUM_GFX_RINGS 1
41 #define GFX9_MEC_HPD_SIZE 2048
42 #define RLCG_UCODE_LOADING_START_ADDRESS 0x00002000L
43 #define RLC_SAVE_RESTORE_ADDR_STARTING_OFFSET 0x00000000L
44 #define GFX9_RLC_FORMAT_DIRECT_REG_LIST_LENGTH 34
46 #define mmPWR_MISC_CNTL_STATUS 0x0183
47 #define mmPWR_MISC_CNTL_STATUS_BASE_IDX 0
48 #define PWR_MISC_CNTL_STATUS__PWR_GFX_RLC_CGPG_EN__SHIFT 0x0
49 #define PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS__SHIFT 0x1
50 #define PWR_MISC_CNTL_STATUS__PWR_GFX_RLC_CGPG_EN_MASK 0x00000001L
51 #define PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS_MASK 0x00000006L
53 MODULE_FIRMWARE("amdgpu/vega10_ce.bin");
54 MODULE_FIRMWARE("amdgpu/vega10_pfp.bin");
55 MODULE_FIRMWARE("amdgpu/vega10_me.bin");
56 MODULE_FIRMWARE("amdgpu/vega10_mec.bin");
57 MODULE_FIRMWARE("amdgpu/vega10_mec2.bin");
58 MODULE_FIRMWARE("amdgpu/vega10_rlc.bin");
60 MODULE_FIRMWARE("amdgpu/raven_ce.bin");
61 MODULE_FIRMWARE("amdgpu/raven_pfp.bin");
62 MODULE_FIRMWARE("amdgpu/raven_me.bin");
63 MODULE_FIRMWARE("amdgpu/raven_mec.bin");
64 MODULE_FIRMWARE("amdgpu/raven_mec2.bin");
65 MODULE_FIRMWARE("amdgpu/raven_rlc.bin");
67 static const struct amdgpu_gds_reg_offset amdgpu_gds_reg_offset[] =
69 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID0_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID0_SIZE),
70 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID0), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID0)},
71 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID1_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID1_SIZE),
72 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID1), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID1)},
73 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID2_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID2_SIZE),
74 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID2), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID2)},
75 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID3_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID3_SIZE),
76 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID3), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID3)},
77 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID4_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID4_SIZE),
78 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID4), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID4)},
79 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID5_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID5_SIZE),
80 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID5), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID5)},
81 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID6_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID6_SIZE),
82 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID6), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID6)},
83 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID7_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID7_SIZE),
84 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID7), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID7)},
85 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID8_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID8_SIZE),
86 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID8), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID8)},
87 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID9_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID9_SIZE),
88 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID9), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID9)},
89 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID10_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID10_SIZE),
90 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID10), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID10)},
91 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID11_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID11_SIZE),
92 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID11), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID11)},
93 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID12_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID12_SIZE),
94 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID12), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID12)},
95 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID13_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID13_SIZE),
96 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID13), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID13)},
97 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID14_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID14_SIZE),
98 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID14), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID14)},
99 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID15_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID15_SIZE),
100 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID15), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID15)}
103 static const u32 golden_settings_gc_9_0[] =
105 SOC15_REG_OFFSET(GC, 0, mmCPC_UTCL1_CNTL), 0x08000000, 0x08000080,
106 SOC15_REG_OFFSET(GC, 0, mmCPF_UTCL1_CNTL), 0x08000000, 0x08000080,
107 SOC15_REG_OFFSET(GC, 0, mmCPG_UTCL1_CNTL), 0x08000000, 0x08000080,
108 SOC15_REG_OFFSET(GC, 0, mmDB_DEBUG2), 0xf00fffff, 0x00000420,
109 SOC15_REG_OFFSET(GC, 0, mmGB_GPU_ID), 0x0000000f, 0x00000000,
110 SOC15_REG_OFFSET(GC, 0, mmIA_UTCL1_CNTL), 0x08000000, 0x08000080,
111 SOC15_REG_OFFSET(GC, 0, mmPA_SC_BINNER_EVENT_CNTL_3), 0x00000003, 0x82400024,
112 SOC15_REG_OFFSET(GC, 0, mmPA_SC_ENHANCE), 0x3fffffff, 0x00000001,
113 SOC15_REG_OFFSET(GC, 0, mmPA_SC_LINE_STIPPLE_STATE), 0x0000ff0f, 0x00000000,
114 SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_UTCL1_CNTL_0), 0x08000000, 0x08000080,
115 SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_UTCL1_CNTL_1), 0x08000000, 0x08000080,
116 SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_UTCL1_CNTL_2), 0x08000000, 0x08000080,
117 SOC15_REG_OFFSET(GC, 0, mmRLC_PREWALKER_UTCL1_CNTL), 0x08000000, 0x08000080,
118 SOC15_REG_OFFSET(GC, 0, mmRLC_SPM_UTCL1_CNTL), 0x08000000, 0x08000080,
119 SOC15_REG_OFFSET(GC, 0, mmSPI_CONFIG_CNTL_1), 0x0000000f, 0x01000107,
120 SOC15_REG_OFFSET(GC, 0, mmTA_CNTL_AUX), 0xfffffeef, 0x010b0000,
121 SOC15_REG_OFFSET(GC, 0, mmTCP_CHAN_STEER_HI), 0xffffffff, 0x4a2c0e68,
122 SOC15_REG_OFFSET(GC, 0, mmTCP_CHAN_STEER_LO), 0xffffffff, 0xb5d3f197,
123 SOC15_REG_OFFSET(GC, 0, mmVGT_CACHE_INVALIDATION), 0x3fff3af3, 0x19200000,
124 SOC15_REG_OFFSET(GC, 0, mmVGT_GS_MAX_WAVE_ID), 0x00000fff, 0x000003ff,
125 SOC15_REG_OFFSET(GC, 0, mmWD_UTCL1_CNTL), 0x08000000, 0x08000080
128 static const u32 golden_settings_gc_9_0_vg10[] =
130 SOC15_REG_OFFSET(GC, 0, mmCB_HW_CONTROL), 0x0000f000, 0x00012107,
131 SOC15_REG_OFFSET(GC, 0, mmCB_HW_CONTROL_3), 0x30000000, 0x10000000,
132 SOC15_REG_OFFSET(GC, 0, mmGB_ADDR_CONFIG), 0xffff77ff, 0x2a114042,
133 SOC15_REG_OFFSET(GC, 0, mmGB_ADDR_CONFIG_READ), 0xffff77ff, 0x2a114042,
134 SOC15_REG_OFFSET(GC, 0, mmPA_SC_ENHANCE_1), 0x00008000, 0x00048000,
135 SOC15_REG_OFFSET(GC, 0, mmRMI_UTCL1_CNTL2), 0x00030000, 0x00020000,
136 SOC15_REG_OFFSET(GC, 0, mmTD_CNTL), 0x00001800, 0x00000800
139 static const u32 golden_settings_gc_9_1[] =
141 SOC15_REG_OFFSET(GC, 0, mmCB_HW_CONTROL), 0xfffdf3cf, 0x00014104,
142 SOC15_REG_OFFSET(GC, 0, mmCPC_UTCL1_CNTL), 0x08000000, 0x08000080,
143 SOC15_REG_OFFSET(GC, 0, mmCPF_UTCL1_CNTL), 0x08000000, 0x08000080,
144 SOC15_REG_OFFSET(GC, 0, mmCPG_UTCL1_CNTL), 0x08000000, 0x08000080,
145 SOC15_REG_OFFSET(GC, 0, mmDB_DEBUG2), 0xf00fffff, 0x00000420,
146 SOC15_REG_OFFSET(GC, 0, mmGB_GPU_ID), 0x0000000f, 0x00000000,
147 SOC15_REG_OFFSET(GC, 0, mmIA_UTCL1_CNTL), 0x08000000, 0x08000080,
148 SOC15_REG_OFFSET(GC, 0, mmPA_SC_BINNER_EVENT_CNTL_3), 0x00000003, 0x82400024,
149 SOC15_REG_OFFSET(GC, 0, mmPA_SC_ENHANCE), 0x3fffffff, 0x00000001,
150 SOC15_REG_OFFSET(GC, 0, mmPA_SC_LINE_STIPPLE_STATE), 0x0000ff0f, 0x00000000,
151 SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_UTCL1_CNTL_0), 0x08000000, 0x08000080,
152 SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_UTCL1_CNTL_1), 0x08000000, 0x08000080,
153 SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_UTCL1_CNTL_2), 0x08000000, 0x08000080,
154 SOC15_REG_OFFSET(GC, 0, mmRLC_PREWALKER_UTCL1_CNTL), 0x08000000, 0x08000080,
155 SOC15_REG_OFFSET(GC, 0, mmRLC_SPM_UTCL1_CNTL), 0x08000000, 0x08000080,
156 SOC15_REG_OFFSET(GC, 0, mmTA_CNTL_AUX), 0xfffffeef, 0x010b0000,
157 SOC15_REG_OFFSET(GC, 0, mmTCP_CHAN_STEER_HI), 0xffffffff, 0x00000000,
158 SOC15_REG_OFFSET(GC, 0, mmTCP_CHAN_STEER_LO), 0xffffffff, 0x00003120,
159 SOC15_REG_OFFSET(GC, 0, mmVGT_CACHE_INVALIDATION), 0x3fff3af3, 0x19200000,
160 SOC15_REG_OFFSET(GC, 0, mmVGT_GS_MAX_WAVE_ID), 0x00000fff, 0x000000ff,
161 SOC15_REG_OFFSET(GC, 0, mmWD_UTCL1_CNTL), 0x08000000, 0x08000080
164 static const u32 golden_settings_gc_9_1_rv1[] =
166 SOC15_REG_OFFSET(GC, 0, mmCB_HW_CONTROL_3), 0x30000000, 0x10000000,
167 SOC15_REG_OFFSET(GC, 0, mmGB_ADDR_CONFIG), 0xffff77ff, 0x24000042,
168 SOC15_REG_OFFSET(GC, 0, mmGB_ADDR_CONFIG_READ), 0xffff77ff, 0x24000042,
169 SOC15_REG_OFFSET(GC, 0, mmPA_SC_ENHANCE_1), 0xffffffff, 0x04048000,
170 SOC15_REG_OFFSET(GC, 0, mmPA_SC_MODE_CNTL_1), 0x06000000, 0x06000000,
171 SOC15_REG_OFFSET(GC, 0, mmRMI_UTCL1_CNTL2), 0x00030000, 0x00020000,
172 SOC15_REG_OFFSET(GC, 0, mmTD_CNTL), 0x01bd9f33, 0x00000800
175 #define VEGA10_GB_ADDR_CONFIG_GOLDEN 0x2a114042
176 #define RAVEN_GB_ADDR_CONFIG_GOLDEN 0x24000042
178 static void gfx_v9_0_set_ring_funcs(struct amdgpu_device *adev);
179 static void gfx_v9_0_set_irq_funcs(struct amdgpu_device *adev);
180 static void gfx_v9_0_set_gds_init(struct amdgpu_device *adev);
181 static void gfx_v9_0_set_rlc_funcs(struct amdgpu_device *adev);
182 static int gfx_v9_0_get_cu_info(struct amdgpu_device *adev,
183 struct amdgpu_cu_info *cu_info);
184 static uint64_t gfx_v9_0_get_gpu_clock_counter(struct amdgpu_device *adev);
185 static void gfx_v9_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, u32 sh_num, u32 instance);
186 static void gfx_v9_0_ring_emit_de_meta(struct amdgpu_ring *ring);
188 static void gfx_v9_0_init_golden_registers(struct amdgpu_device *adev)
190 switch (adev->asic_type) {
192 amdgpu_program_register_sequence(adev,
193 golden_settings_gc_9_0,
194 (const u32)ARRAY_SIZE(golden_settings_gc_9_0));
195 amdgpu_program_register_sequence(adev,
196 golden_settings_gc_9_0_vg10,
197 (const u32)ARRAY_SIZE(golden_settings_gc_9_0_vg10));
200 amdgpu_program_register_sequence(adev,
201 golden_settings_gc_9_1,
202 (const u32)ARRAY_SIZE(golden_settings_gc_9_1));
203 amdgpu_program_register_sequence(adev,
204 golden_settings_gc_9_1_rv1,
205 (const u32)ARRAY_SIZE(golden_settings_gc_9_1_rv1));
212 static void gfx_v9_0_scratch_init(struct amdgpu_device *adev)
214 adev->gfx.scratch.num_reg = 7;
215 adev->gfx.scratch.reg_base = SOC15_REG_OFFSET(GC, 0, mmSCRATCH_REG0);
216 adev->gfx.scratch.free_mask = (1u << adev->gfx.scratch.num_reg) - 1;
219 static void gfx_v9_0_write_data_to_reg(struct amdgpu_ring *ring, int eng_sel,
220 bool wc, uint32_t reg, uint32_t val)
222 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
223 amdgpu_ring_write(ring, WRITE_DATA_ENGINE_SEL(eng_sel) |
224 WRITE_DATA_DST_SEL(0) |
225 (wc ? WR_CONFIRM : 0));
226 amdgpu_ring_write(ring, reg);
227 amdgpu_ring_write(ring, 0);
228 amdgpu_ring_write(ring, val);
231 static void gfx_v9_0_wait_reg_mem(struct amdgpu_ring *ring, int eng_sel,
232 int mem_space, int opt, uint32_t addr0,
233 uint32_t addr1, uint32_t ref, uint32_t mask,
236 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
237 amdgpu_ring_write(ring,
238 /* memory (1) or register (0) */
239 (WAIT_REG_MEM_MEM_SPACE(mem_space) |
240 WAIT_REG_MEM_OPERATION(opt) | /* wait */
241 WAIT_REG_MEM_FUNCTION(3) | /* equal */
242 WAIT_REG_MEM_ENGINE(eng_sel)));
245 BUG_ON(addr0 & 0x3); /* Dword align */
246 amdgpu_ring_write(ring, addr0);
247 amdgpu_ring_write(ring, addr1);
248 amdgpu_ring_write(ring, ref);
249 amdgpu_ring_write(ring, mask);
250 amdgpu_ring_write(ring, inv); /* poll interval */
253 static int gfx_v9_0_ring_test_ring(struct amdgpu_ring *ring)
255 struct amdgpu_device *adev = ring->adev;
261 r = amdgpu_gfx_scratch_get(adev, &scratch);
263 DRM_ERROR("amdgpu: cp failed to get scratch reg (%d).\n", r);
266 WREG32(scratch, 0xCAFEDEAD);
267 r = amdgpu_ring_alloc(ring, 3);
269 DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n",
271 amdgpu_gfx_scratch_free(adev, scratch);
274 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1));
275 amdgpu_ring_write(ring, (scratch - PACKET3_SET_UCONFIG_REG_START));
276 amdgpu_ring_write(ring, 0xDEADBEEF);
277 amdgpu_ring_commit(ring);
279 for (i = 0; i < adev->usec_timeout; i++) {
280 tmp = RREG32(scratch);
281 if (tmp == 0xDEADBEEF)
285 if (i < adev->usec_timeout) {
286 DRM_INFO("ring test on %d succeeded in %d usecs\n",
289 DRM_ERROR("amdgpu: ring %d test failed (scratch(0x%04X)=0x%08X)\n",
290 ring->idx, scratch, tmp);
293 amdgpu_gfx_scratch_free(adev, scratch);
297 static int gfx_v9_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
299 struct amdgpu_device *adev = ring->adev;
301 struct dma_fence *f = NULL;
306 r = amdgpu_gfx_scratch_get(adev, &scratch);
308 DRM_ERROR("amdgpu: failed to get scratch reg (%ld).\n", r);
311 WREG32(scratch, 0xCAFEDEAD);
312 memset(&ib, 0, sizeof(ib));
313 r = amdgpu_ib_get(adev, NULL, 256, &ib);
315 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
318 ib.ptr[0] = PACKET3(PACKET3_SET_UCONFIG_REG, 1);
319 ib.ptr[1] = ((scratch - PACKET3_SET_UCONFIG_REG_START));
320 ib.ptr[2] = 0xDEADBEEF;
323 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
327 r = dma_fence_wait_timeout(f, false, timeout);
329 DRM_ERROR("amdgpu: IB test timed out.\n");
333 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
336 tmp = RREG32(scratch);
337 if (tmp == 0xDEADBEEF) {
338 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
341 DRM_ERROR("amdgpu: ib test failed (scratch(0x%04X)=0x%08X)\n",
346 amdgpu_ib_free(adev, &ib, NULL);
349 amdgpu_gfx_scratch_free(adev, scratch);
353 static int gfx_v9_0_init_microcode(struct amdgpu_device *adev)
355 const char *chip_name;
358 struct amdgpu_firmware_info *info = NULL;
359 const struct common_firmware_header *header = NULL;
360 const struct gfx_firmware_header_v1_0 *cp_hdr;
361 const struct rlc_firmware_header_v2_0 *rlc_hdr;
362 unsigned int *tmp = NULL;
367 switch (adev->asic_type) {
369 chip_name = "vega10";
378 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_pfp.bin", chip_name);
379 err = request_firmware(&adev->gfx.pfp_fw, fw_name, adev->dev);
382 err = amdgpu_ucode_validate(adev->gfx.pfp_fw);
385 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.pfp_fw->data;
386 adev->gfx.pfp_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
387 adev->gfx.pfp_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
389 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_me.bin", chip_name);
390 err = request_firmware(&adev->gfx.me_fw, fw_name, adev->dev);
393 err = amdgpu_ucode_validate(adev->gfx.me_fw);
396 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.me_fw->data;
397 adev->gfx.me_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
398 adev->gfx.me_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
400 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_ce.bin", chip_name);
401 err = request_firmware(&adev->gfx.ce_fw, fw_name, adev->dev);
404 err = amdgpu_ucode_validate(adev->gfx.ce_fw);
407 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.ce_fw->data;
408 adev->gfx.ce_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
409 adev->gfx.ce_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
411 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_rlc.bin", chip_name);
412 err = request_firmware(&adev->gfx.rlc_fw, fw_name, adev->dev);
415 err = amdgpu_ucode_validate(adev->gfx.rlc_fw);
416 rlc_hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data;
417 adev->gfx.rlc_fw_version = le32_to_cpu(rlc_hdr->header.ucode_version);
418 adev->gfx.rlc_feature_version = le32_to_cpu(rlc_hdr->ucode_feature_version);
419 adev->gfx.rlc.save_and_restore_offset =
420 le32_to_cpu(rlc_hdr->save_and_restore_offset);
421 adev->gfx.rlc.clear_state_descriptor_offset =
422 le32_to_cpu(rlc_hdr->clear_state_descriptor_offset);
423 adev->gfx.rlc.avail_scratch_ram_locations =
424 le32_to_cpu(rlc_hdr->avail_scratch_ram_locations);
425 adev->gfx.rlc.reg_restore_list_size =
426 le32_to_cpu(rlc_hdr->reg_restore_list_size);
427 adev->gfx.rlc.reg_list_format_start =
428 le32_to_cpu(rlc_hdr->reg_list_format_start);
429 adev->gfx.rlc.reg_list_format_separate_start =
430 le32_to_cpu(rlc_hdr->reg_list_format_separate_start);
431 adev->gfx.rlc.starting_offsets_start =
432 le32_to_cpu(rlc_hdr->starting_offsets_start);
433 adev->gfx.rlc.reg_list_format_size_bytes =
434 le32_to_cpu(rlc_hdr->reg_list_format_size_bytes);
435 adev->gfx.rlc.reg_list_size_bytes =
436 le32_to_cpu(rlc_hdr->reg_list_size_bytes);
437 adev->gfx.rlc.register_list_format =
438 kmalloc(adev->gfx.rlc.reg_list_format_size_bytes +
439 adev->gfx.rlc.reg_list_size_bytes, GFP_KERNEL);
440 if (!adev->gfx.rlc.register_list_format) {
445 tmp = (unsigned int *)((uintptr_t)rlc_hdr +
446 le32_to_cpu(rlc_hdr->reg_list_format_array_offset_bytes));
447 for (i = 0 ; i < (rlc_hdr->reg_list_format_size_bytes >> 2); i++)
448 adev->gfx.rlc.register_list_format[i] = le32_to_cpu(tmp[i]);
450 adev->gfx.rlc.register_restore = adev->gfx.rlc.register_list_format + i;
452 tmp = (unsigned int *)((uintptr_t)rlc_hdr +
453 le32_to_cpu(rlc_hdr->reg_list_array_offset_bytes));
454 for (i = 0 ; i < (rlc_hdr->reg_list_size_bytes >> 2); i++)
455 adev->gfx.rlc.register_restore[i] = le32_to_cpu(tmp[i]);
457 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mec.bin", chip_name);
458 err = request_firmware(&adev->gfx.mec_fw, fw_name, adev->dev);
461 err = amdgpu_ucode_validate(adev->gfx.mec_fw);
464 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
465 adev->gfx.mec_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
466 adev->gfx.mec_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
469 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mec2.bin", chip_name);
470 err = request_firmware(&adev->gfx.mec2_fw, fw_name, adev->dev);
472 err = amdgpu_ucode_validate(adev->gfx.mec2_fw);
475 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
476 adev->gfx.mec2_fw->data;
477 adev->gfx.mec2_fw_version =
478 le32_to_cpu(cp_hdr->header.ucode_version);
479 adev->gfx.mec2_feature_version =
480 le32_to_cpu(cp_hdr->ucode_feature_version);
483 adev->gfx.mec2_fw = NULL;
486 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
487 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_PFP];
488 info->ucode_id = AMDGPU_UCODE_ID_CP_PFP;
489 info->fw = adev->gfx.pfp_fw;
490 header = (const struct common_firmware_header *)info->fw->data;
491 adev->firmware.fw_size +=
492 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
494 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_ME];
495 info->ucode_id = AMDGPU_UCODE_ID_CP_ME;
496 info->fw = adev->gfx.me_fw;
497 header = (const struct common_firmware_header *)info->fw->data;
498 adev->firmware.fw_size +=
499 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
501 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_CE];
502 info->ucode_id = AMDGPU_UCODE_ID_CP_CE;
503 info->fw = adev->gfx.ce_fw;
504 header = (const struct common_firmware_header *)info->fw->data;
505 adev->firmware.fw_size +=
506 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
508 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_RLC_G];
509 info->ucode_id = AMDGPU_UCODE_ID_RLC_G;
510 info->fw = adev->gfx.rlc_fw;
511 header = (const struct common_firmware_header *)info->fw->data;
512 adev->firmware.fw_size +=
513 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
515 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_MEC1];
516 info->ucode_id = AMDGPU_UCODE_ID_CP_MEC1;
517 info->fw = adev->gfx.mec_fw;
518 header = (const struct common_firmware_header *)info->fw->data;
519 cp_hdr = (const struct gfx_firmware_header_v1_0 *)info->fw->data;
520 adev->firmware.fw_size +=
521 ALIGN(le32_to_cpu(header->ucode_size_bytes) - le32_to_cpu(cp_hdr->jt_size) * 4, PAGE_SIZE);
523 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_MEC1_JT];
524 info->ucode_id = AMDGPU_UCODE_ID_CP_MEC1_JT;
525 info->fw = adev->gfx.mec_fw;
526 adev->firmware.fw_size +=
527 ALIGN(le32_to_cpu(cp_hdr->jt_size) * 4, PAGE_SIZE);
529 if (adev->gfx.mec2_fw) {
530 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_MEC2];
531 info->ucode_id = AMDGPU_UCODE_ID_CP_MEC2;
532 info->fw = adev->gfx.mec2_fw;
533 header = (const struct common_firmware_header *)info->fw->data;
534 cp_hdr = (const struct gfx_firmware_header_v1_0 *)info->fw->data;
535 adev->firmware.fw_size +=
536 ALIGN(le32_to_cpu(header->ucode_size_bytes) - le32_to_cpu(cp_hdr->jt_size) * 4, PAGE_SIZE);
537 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_MEC2_JT];
538 info->ucode_id = AMDGPU_UCODE_ID_CP_MEC2_JT;
539 info->fw = adev->gfx.mec2_fw;
540 adev->firmware.fw_size +=
541 ALIGN(le32_to_cpu(cp_hdr->jt_size) * 4, PAGE_SIZE);
549 "gfx9: Failed to load firmware \"%s\"\n",
551 release_firmware(adev->gfx.pfp_fw);
552 adev->gfx.pfp_fw = NULL;
553 release_firmware(adev->gfx.me_fw);
554 adev->gfx.me_fw = NULL;
555 release_firmware(adev->gfx.ce_fw);
556 adev->gfx.ce_fw = NULL;
557 release_firmware(adev->gfx.rlc_fw);
558 adev->gfx.rlc_fw = NULL;
559 release_firmware(adev->gfx.mec_fw);
560 adev->gfx.mec_fw = NULL;
561 release_firmware(adev->gfx.mec2_fw);
562 adev->gfx.mec2_fw = NULL;
567 static u32 gfx_v9_0_get_csb_size(struct amdgpu_device *adev)
570 const struct cs_section_def *sect = NULL;
571 const struct cs_extent_def *ext = NULL;
573 /* begin clear state */
575 /* context control state */
578 for (sect = gfx9_cs_data; sect->section != NULL; ++sect) {
579 for (ext = sect->section; ext->extent != NULL; ++ext) {
580 if (sect->id == SECT_CONTEXT)
581 count += 2 + ext->reg_count;
587 /* end clear state */
595 static void gfx_v9_0_get_csb_buffer(struct amdgpu_device *adev,
596 volatile u32 *buffer)
599 const struct cs_section_def *sect = NULL;
600 const struct cs_extent_def *ext = NULL;
602 if (adev->gfx.rlc.cs_data == NULL)
607 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
608 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
610 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1));
611 buffer[count++] = cpu_to_le32(0x80000000);
612 buffer[count++] = cpu_to_le32(0x80000000);
614 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) {
615 for (ext = sect->section; ext->extent != NULL; ++ext) {
616 if (sect->id == SECT_CONTEXT) {
618 cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count));
619 buffer[count++] = cpu_to_le32(ext->reg_index -
620 PACKET3_SET_CONTEXT_REG_START);
621 for (i = 0; i < ext->reg_count; i++)
622 buffer[count++] = cpu_to_le32(ext->extent[i]);
629 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0));
630 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE);
632 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0));
633 buffer[count++] = cpu_to_le32(0);
636 static void gfx_v9_0_init_lbpw(struct amdgpu_device *adev)
640 /* set mmRLC_LB_THR_CONFIG_1/2/3/4 */
641 WREG32_SOC15(GC, 0, mmRLC_LB_THR_CONFIG_1, 0x0000007F);
642 WREG32_SOC15(GC, 0, mmRLC_LB_THR_CONFIG_2, 0x0333A5A7);
643 WREG32_SOC15(GC, 0, mmRLC_LB_THR_CONFIG_3, 0x00000077);
644 WREG32_SOC15(GC, 0, mmRLC_LB_THR_CONFIG_4, (0x30 | 0x40 << 8 | 0x02FA << 16));
646 /* set mmRLC_LB_CNTR_INIT = 0x0000_0000 */
647 WREG32_SOC15(GC, 0, mmRLC_LB_CNTR_INIT, 0x00000000);
649 /* set mmRLC_LB_CNTR_MAX = 0x0000_0500 */
650 WREG32_SOC15(GC, 0, mmRLC_LB_CNTR_MAX, 0x00000500);
652 mutex_lock(&adev->grbm_idx_mutex);
653 /* set mmRLC_LB_INIT_CU_MASK thru broadcast mode to enable all SE/SH*/
654 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
655 WREG32_SOC15(GC, 0, mmRLC_LB_INIT_CU_MASK, 0xffffffff);
657 /* set mmRLC_LB_PARAMS = 0x003F_1006 */
658 data = REG_SET_FIELD(0, RLC_LB_PARAMS, FIFO_SAMPLES, 0x0003);
659 data |= REG_SET_FIELD(data, RLC_LB_PARAMS, PG_IDLE_SAMPLES, 0x0010);
660 data |= REG_SET_FIELD(data, RLC_LB_PARAMS, PG_IDLE_SAMPLE_INTERVAL, 0x033F);
661 WREG32_SOC15(GC, 0, mmRLC_LB_PARAMS, data);
663 /* set mmRLC_GPM_GENERAL_7[31-16] = 0x00C0 */
664 data = RREG32_SOC15(GC, 0, mmRLC_GPM_GENERAL_7);
667 WREG32_SOC15(GC, 0, mmRLC_GPM_GENERAL_7, data);
669 /* set RLC_LB_ALWAYS_ACTIVE_CU_MASK = 0xFFF */
670 WREG32_SOC15(GC, 0, mmRLC_LB_ALWAYS_ACTIVE_CU_MASK, 0xFFF);
672 /* set RLC_LB_CNTL = 0x8000_0095, 31 bit is reserved,
673 * but used for RLC_LB_CNTL configuration */
674 data = RLC_LB_CNTL__LB_CNT_SPIM_ACTIVE_MASK;
675 data |= REG_SET_FIELD(data, RLC_LB_CNTL, CU_MASK_USED_OFF_HYST, 0x09);
676 data |= REG_SET_FIELD(data, RLC_LB_CNTL, RESERVED, 0x80000);
677 WREG32_SOC15(GC, 0, mmRLC_LB_CNTL, data);
678 mutex_unlock(&adev->grbm_idx_mutex);
681 static void gfx_v9_0_enable_lbpw(struct amdgpu_device *adev, bool enable)
683 WREG32_FIELD15(GC, 0, RLC_LB_CNTL, LOAD_BALANCE_ENABLE, enable ? 1 : 0);
686 static void rv_init_cp_jump_table(struct amdgpu_device *adev)
688 const __le32 *fw_data;
689 volatile u32 *dst_ptr;
690 int me, i, max_me = 5;
692 u32 table_offset, table_size;
694 /* write the cp table buffer */
695 dst_ptr = adev->gfx.rlc.cp_table_ptr;
696 for (me = 0; me < max_me; me++) {
698 const struct gfx_firmware_header_v1_0 *hdr =
699 (const struct gfx_firmware_header_v1_0 *)adev->gfx.ce_fw->data;
700 fw_data = (const __le32 *)
701 (adev->gfx.ce_fw->data +
702 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
703 table_offset = le32_to_cpu(hdr->jt_offset);
704 table_size = le32_to_cpu(hdr->jt_size);
705 } else if (me == 1) {
706 const struct gfx_firmware_header_v1_0 *hdr =
707 (const struct gfx_firmware_header_v1_0 *)adev->gfx.pfp_fw->data;
708 fw_data = (const __le32 *)
709 (adev->gfx.pfp_fw->data +
710 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
711 table_offset = le32_to_cpu(hdr->jt_offset);
712 table_size = le32_to_cpu(hdr->jt_size);
713 } else if (me == 2) {
714 const struct gfx_firmware_header_v1_0 *hdr =
715 (const struct gfx_firmware_header_v1_0 *)adev->gfx.me_fw->data;
716 fw_data = (const __le32 *)
717 (adev->gfx.me_fw->data +
718 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
719 table_offset = le32_to_cpu(hdr->jt_offset);
720 table_size = le32_to_cpu(hdr->jt_size);
721 } else if (me == 3) {
722 const struct gfx_firmware_header_v1_0 *hdr =
723 (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
724 fw_data = (const __le32 *)
725 (adev->gfx.mec_fw->data +
726 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
727 table_offset = le32_to_cpu(hdr->jt_offset);
728 table_size = le32_to_cpu(hdr->jt_size);
729 } else if (me == 4) {
730 const struct gfx_firmware_header_v1_0 *hdr =
731 (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec2_fw->data;
732 fw_data = (const __le32 *)
733 (adev->gfx.mec2_fw->data +
734 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
735 table_offset = le32_to_cpu(hdr->jt_offset);
736 table_size = le32_to_cpu(hdr->jt_size);
739 for (i = 0; i < table_size; i ++) {
740 dst_ptr[bo_offset + i] =
741 cpu_to_le32(le32_to_cpu(fw_data[table_offset + i]));
744 bo_offset += table_size;
748 static void gfx_v9_0_rlc_fini(struct amdgpu_device *adev)
750 /* clear state block */
751 amdgpu_bo_free_kernel(&adev->gfx.rlc.clear_state_obj,
752 &adev->gfx.rlc.clear_state_gpu_addr,
753 (void **)&adev->gfx.rlc.cs_ptr);
755 /* jump table block */
756 amdgpu_bo_free_kernel(&adev->gfx.rlc.cp_table_obj,
757 &adev->gfx.rlc.cp_table_gpu_addr,
758 (void **)&adev->gfx.rlc.cp_table_ptr);
761 static int gfx_v9_0_rlc_init(struct amdgpu_device *adev)
763 volatile u32 *dst_ptr;
765 const struct cs_section_def *cs_data;
768 adev->gfx.rlc.cs_data = gfx9_cs_data;
770 cs_data = adev->gfx.rlc.cs_data;
773 /* clear state block */
774 adev->gfx.rlc.clear_state_size = dws = gfx_v9_0_get_csb_size(adev);
775 if (adev->gfx.rlc.clear_state_obj == NULL) {
776 r = amdgpu_bo_create_kernel(adev, dws * 4, PAGE_SIZE,
777 AMDGPU_GEM_DOMAIN_VRAM,
778 &adev->gfx.rlc.clear_state_obj,
779 &adev->gfx.rlc.clear_state_gpu_addr,
780 (void **)&adev->gfx.rlc.cs_ptr);
783 "(%d) failed to create rlc csb bo\n", r);
784 gfx_v9_0_rlc_fini(adev);
788 /* set up the cs buffer */
789 dst_ptr = adev->gfx.rlc.cs_ptr;
790 gfx_v9_0_get_csb_buffer(adev, dst_ptr);
791 amdgpu_bo_kunmap(adev->gfx.rlc.clear_state_obj);
792 amdgpu_bo_unreserve(adev->gfx.rlc.clear_state_obj);
795 if (adev->asic_type == CHIP_RAVEN) {
796 /* TODO: double check the cp_table_size for RV */
797 adev->gfx.rlc.cp_table_size = ALIGN(96 * 5 * 4, 2048) + (64 * 1024); /* JT + GDS */
798 if (adev->gfx.rlc.cp_table_obj == NULL) {
799 r = amdgpu_bo_create_kernel(adev, adev->gfx.rlc.cp_table_size,
800 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
801 &adev->gfx.rlc.cp_table_obj,
802 &adev->gfx.rlc.cp_table_gpu_addr,
803 (void **)&adev->gfx.rlc.cp_table_ptr);
806 "(%d) failed to create cp table bo\n", r);
807 gfx_v9_0_rlc_fini(adev);
812 rv_init_cp_jump_table(adev);
813 amdgpu_bo_kunmap(adev->gfx.rlc.cp_table_obj);
814 amdgpu_bo_unreserve(adev->gfx.rlc.cp_table_obj);
816 gfx_v9_0_init_lbpw(adev);
822 static void gfx_v9_0_mec_fini(struct amdgpu_device *adev)
826 if (adev->gfx.mec.hpd_eop_obj) {
827 r = amdgpu_bo_reserve(adev->gfx.mec.hpd_eop_obj, true);
828 if (unlikely(r != 0))
829 dev_warn(adev->dev, "(%d) reserve HPD EOP bo failed\n", r);
830 amdgpu_bo_unpin(adev->gfx.mec.hpd_eop_obj);
831 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
833 amdgpu_bo_unref(&adev->gfx.mec.hpd_eop_obj);
834 adev->gfx.mec.hpd_eop_obj = NULL;
836 if (adev->gfx.mec.mec_fw_obj) {
837 r = amdgpu_bo_reserve(adev->gfx.mec.mec_fw_obj, true);
838 if (unlikely(r != 0))
839 dev_warn(adev->dev, "(%d) reserve mec firmware bo failed\n", r);
840 amdgpu_bo_unpin(adev->gfx.mec.mec_fw_obj);
841 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj);
843 amdgpu_bo_unref(&adev->gfx.mec.mec_fw_obj);
844 adev->gfx.mec.mec_fw_obj = NULL;
848 static int gfx_v9_0_mec_init(struct amdgpu_device *adev)
852 const __le32 *fw_data;
857 const struct gfx_firmware_header_v1_0 *mec_hdr;
859 bitmap_zero(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
861 /* take ownership of the relevant compute queues */
862 amdgpu_gfx_compute_queue_acquire(adev);
863 mec_hpd_size = adev->gfx.num_compute_rings * GFX9_MEC_HPD_SIZE;
865 if (adev->gfx.mec.hpd_eop_obj == NULL) {
866 r = amdgpu_bo_create(adev,
869 AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL,
870 &adev->gfx.mec.hpd_eop_obj);
872 dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
877 r = amdgpu_bo_reserve(adev->gfx.mec.hpd_eop_obj, false);
878 if (unlikely(r != 0)) {
879 gfx_v9_0_mec_fini(adev);
882 r = amdgpu_bo_pin(adev->gfx.mec.hpd_eop_obj, AMDGPU_GEM_DOMAIN_GTT,
883 &adev->gfx.mec.hpd_eop_gpu_addr);
885 dev_warn(adev->dev, "(%d) pin HDP EOP bo failed\n", r);
886 gfx_v9_0_mec_fini(adev);
889 r = amdgpu_bo_kmap(adev->gfx.mec.hpd_eop_obj, (void **)&hpd);
891 dev_warn(adev->dev, "(%d) map HDP EOP bo failed\n", r);
892 gfx_v9_0_mec_fini(adev);
896 memset(hpd, 0, adev->gfx.mec.hpd_eop_obj->tbo.mem.size);
898 amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
899 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
901 mec_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
903 fw_data = (const __le32 *)
904 (adev->gfx.mec_fw->data +
905 le32_to_cpu(mec_hdr->header.ucode_array_offset_bytes));
906 fw_size = le32_to_cpu(mec_hdr->header.ucode_size_bytes) / 4;
908 if (adev->gfx.mec.mec_fw_obj == NULL) {
909 r = amdgpu_bo_create(adev,
910 mec_hdr->header.ucode_size_bytes,
912 AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL,
913 &adev->gfx.mec.mec_fw_obj);
915 dev_warn(adev->dev, "(%d) create mec firmware bo failed\n", r);
920 r = amdgpu_bo_reserve(adev->gfx.mec.mec_fw_obj, false);
921 if (unlikely(r != 0)) {
922 gfx_v9_0_mec_fini(adev);
925 r = amdgpu_bo_pin(adev->gfx.mec.mec_fw_obj, AMDGPU_GEM_DOMAIN_GTT,
926 &adev->gfx.mec.mec_fw_gpu_addr);
928 dev_warn(adev->dev, "(%d) pin mec firmware bo failed\n", r);
929 gfx_v9_0_mec_fini(adev);
932 r = amdgpu_bo_kmap(adev->gfx.mec.mec_fw_obj, (void **)&fw);
934 dev_warn(adev->dev, "(%d) map firmware bo failed\n", r);
935 gfx_v9_0_mec_fini(adev);
938 memcpy(fw, fw_data, fw_size);
940 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_obj);
941 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj);
947 static uint32_t wave_read_ind(struct amdgpu_device *adev, uint32_t simd, uint32_t wave, uint32_t address)
949 WREG32_SOC15(GC, 0, mmSQ_IND_INDEX,
950 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) |
951 (simd << SQ_IND_INDEX__SIMD_ID__SHIFT) |
952 (address << SQ_IND_INDEX__INDEX__SHIFT) |
953 (SQ_IND_INDEX__FORCE_READ_MASK));
954 return RREG32_SOC15(GC, 0, mmSQ_IND_DATA);
957 static void wave_read_regs(struct amdgpu_device *adev, uint32_t simd,
958 uint32_t wave, uint32_t thread,
959 uint32_t regno, uint32_t num, uint32_t *out)
961 WREG32_SOC15(GC, 0, mmSQ_IND_INDEX,
962 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) |
963 (simd << SQ_IND_INDEX__SIMD_ID__SHIFT) |
964 (regno << SQ_IND_INDEX__INDEX__SHIFT) |
965 (thread << SQ_IND_INDEX__THREAD_ID__SHIFT) |
966 (SQ_IND_INDEX__FORCE_READ_MASK) |
967 (SQ_IND_INDEX__AUTO_INCR_MASK));
969 *(out++) = RREG32_SOC15(GC, 0, mmSQ_IND_DATA);
972 static void gfx_v9_0_read_wave_data(struct amdgpu_device *adev, uint32_t simd, uint32_t wave, uint32_t *dst, int *no_fields)
974 /* type 1 wave data */
975 dst[(*no_fields)++] = 1;
976 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_STATUS);
977 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_PC_LO);
978 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_PC_HI);
979 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_EXEC_LO);
980 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_EXEC_HI);
981 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_HW_ID);
982 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_INST_DW0);
983 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_INST_DW1);
984 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_GPR_ALLOC);
985 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_LDS_ALLOC);
986 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_TRAPSTS);
987 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_IB_STS);
988 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_IB_DBG0);
989 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_M0);
992 static void gfx_v9_0_read_wave_sgprs(struct amdgpu_device *adev, uint32_t simd,
993 uint32_t wave, uint32_t start,
994 uint32_t size, uint32_t *dst)
998 start + SQIND_WAVE_SGPRS_OFFSET, size, dst);
1002 static const struct amdgpu_gfx_funcs gfx_v9_0_gfx_funcs = {
1003 .get_gpu_clock_counter = &gfx_v9_0_get_gpu_clock_counter,
1004 .select_se_sh = &gfx_v9_0_select_se_sh,
1005 .read_wave_data = &gfx_v9_0_read_wave_data,
1006 .read_wave_sgprs = &gfx_v9_0_read_wave_sgprs,
1009 static void gfx_v9_0_gpu_early_init(struct amdgpu_device *adev)
1013 adev->gfx.funcs = &gfx_v9_0_gfx_funcs;
1015 switch (adev->asic_type) {
1017 adev->gfx.config.max_hw_contexts = 8;
1018 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
1019 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
1020 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
1021 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x4C0;
1022 gb_addr_config = VEGA10_GB_ADDR_CONFIG_GOLDEN;
1025 adev->gfx.config.max_hw_contexts = 8;
1026 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
1027 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
1028 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
1029 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x4C0;
1030 gb_addr_config = RAVEN_GB_ADDR_CONFIG_GOLDEN;
1037 adev->gfx.config.gb_addr_config = gb_addr_config;
1039 adev->gfx.config.gb_addr_config_fields.num_pipes = 1 <<
1041 adev->gfx.config.gb_addr_config,
1045 adev->gfx.config.max_tile_pipes =
1046 adev->gfx.config.gb_addr_config_fields.num_pipes;
1048 adev->gfx.config.gb_addr_config_fields.num_banks = 1 <<
1050 adev->gfx.config.gb_addr_config,
1053 adev->gfx.config.gb_addr_config_fields.max_compress_frags = 1 <<
1055 adev->gfx.config.gb_addr_config,
1057 MAX_COMPRESSED_FRAGS);
1058 adev->gfx.config.gb_addr_config_fields.num_rb_per_se = 1 <<
1060 adev->gfx.config.gb_addr_config,
1063 adev->gfx.config.gb_addr_config_fields.num_se = 1 <<
1065 adev->gfx.config.gb_addr_config,
1067 NUM_SHADER_ENGINES);
1068 adev->gfx.config.gb_addr_config_fields.pipe_interleave_size = 1 << (8 +
1070 adev->gfx.config.gb_addr_config,
1072 PIPE_INTERLEAVE_SIZE));
1075 static int gfx_v9_0_ngg_create_buf(struct amdgpu_device *adev,
1076 struct amdgpu_ngg_buf *ngg_buf,
1078 int default_size_se)
1083 dev_err(adev->dev, "Buffer size is invalid: %d\n", size_se);
1086 size_se = size_se ? size_se : default_size_se;
1088 ngg_buf->size = size_se * adev->gfx.config.max_shader_engines;
1089 r = amdgpu_bo_create_kernel(adev, ngg_buf->size,
1090 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
1095 dev_err(adev->dev, "(%d) failed to create NGG buffer\n", r);
1098 ngg_buf->bo_size = amdgpu_bo_size(ngg_buf->bo);
1103 static int gfx_v9_0_ngg_fini(struct amdgpu_device *adev)
1107 for (i = 0; i < NGG_BUF_MAX; i++)
1108 amdgpu_bo_free_kernel(&adev->gfx.ngg.buf[i].bo,
1109 &adev->gfx.ngg.buf[i].gpu_addr,
1112 memset(&adev->gfx.ngg.buf[0], 0,
1113 sizeof(struct amdgpu_ngg_buf) * NGG_BUF_MAX);
1115 adev->gfx.ngg.init = false;
1120 static int gfx_v9_0_ngg_init(struct amdgpu_device *adev)
1124 if (!amdgpu_ngg || adev->gfx.ngg.init == true)
1127 /* GDS reserve memory: 64 bytes alignment */
1128 adev->gfx.ngg.gds_reserve_size = ALIGN(5 * 4, 0x40);
1129 adev->gds.mem.total_size -= adev->gfx.ngg.gds_reserve_size;
1130 adev->gds.mem.gfx_partition_size -= adev->gfx.ngg.gds_reserve_size;
1131 adev->gfx.ngg.gds_reserve_addr = amdgpu_gds_reg_offset[0].mem_base;
1132 adev->gfx.ngg.gds_reserve_addr += adev->gds.mem.gfx_partition_size;
1134 /* Primitive Buffer */
1135 r = gfx_v9_0_ngg_create_buf(adev, &adev->gfx.ngg.buf[NGG_PRIM],
1136 amdgpu_prim_buf_per_se,
1139 dev_err(adev->dev, "Failed to create Primitive Buffer\n");
1143 /* Position Buffer */
1144 r = gfx_v9_0_ngg_create_buf(adev, &adev->gfx.ngg.buf[NGG_POS],
1145 amdgpu_pos_buf_per_se,
1148 dev_err(adev->dev, "Failed to create Position Buffer\n");
1152 /* Control Sideband */
1153 r = gfx_v9_0_ngg_create_buf(adev, &adev->gfx.ngg.buf[NGG_CNTL],
1154 amdgpu_cntl_sb_buf_per_se,
1157 dev_err(adev->dev, "Failed to create Control Sideband Buffer\n");
1161 /* Parameter Cache, not created by default */
1162 if (amdgpu_param_buf_per_se <= 0)
1165 r = gfx_v9_0_ngg_create_buf(adev, &adev->gfx.ngg.buf[NGG_PARAM],
1166 amdgpu_param_buf_per_se,
1169 dev_err(adev->dev, "Failed to create Parameter Cache\n");
1174 adev->gfx.ngg.init = true;
1177 gfx_v9_0_ngg_fini(adev);
1181 static int gfx_v9_0_ngg_en(struct amdgpu_device *adev)
1183 struct amdgpu_ring *ring = &adev->gfx.gfx_ring[0];
1192 /* Program buffer size */
1194 size = adev->gfx.ngg.buf[NGG_PRIM].size / 256;
1195 data = REG_SET_FIELD(data, WD_BUF_RESOURCE_1, INDEX_BUF_SIZE, size);
1197 size = adev->gfx.ngg.buf[NGG_POS].size / 256;
1198 data = REG_SET_FIELD(data, WD_BUF_RESOURCE_1, POS_BUF_SIZE, size);
1200 WREG32_SOC15(GC, 0, mmWD_BUF_RESOURCE_1, data);
1203 size = adev->gfx.ngg.buf[NGG_CNTL].size / 256;
1204 data = REG_SET_FIELD(data, WD_BUF_RESOURCE_2, CNTL_SB_BUF_SIZE, size);
1206 size = adev->gfx.ngg.buf[NGG_PARAM].size / 1024;
1207 data = REG_SET_FIELD(data, WD_BUF_RESOURCE_2, PARAM_BUF_SIZE, size);
1209 WREG32_SOC15(GC, 0, mmWD_BUF_RESOURCE_2, data);
1211 /* Program buffer base address */
1212 base = lower_32_bits(adev->gfx.ngg.buf[NGG_PRIM].gpu_addr);
1213 data = REG_SET_FIELD(0, WD_INDEX_BUF_BASE, BASE, base);
1214 WREG32_SOC15(GC, 0, mmWD_INDEX_BUF_BASE, data);
1216 base = upper_32_bits(adev->gfx.ngg.buf[NGG_PRIM].gpu_addr);
1217 data = REG_SET_FIELD(0, WD_INDEX_BUF_BASE_HI, BASE_HI, base);
1218 WREG32_SOC15(GC, 0, mmWD_INDEX_BUF_BASE_HI, data);
1220 base = lower_32_bits(adev->gfx.ngg.buf[NGG_POS].gpu_addr);
1221 data = REG_SET_FIELD(0, WD_POS_BUF_BASE, BASE, base);
1222 WREG32_SOC15(GC, 0, mmWD_POS_BUF_BASE, data);
1224 base = upper_32_bits(adev->gfx.ngg.buf[NGG_POS].gpu_addr);
1225 data = REG_SET_FIELD(0, WD_POS_BUF_BASE_HI, BASE_HI, base);
1226 WREG32_SOC15(GC, 0, mmWD_POS_BUF_BASE_HI, data);
1228 base = lower_32_bits(adev->gfx.ngg.buf[NGG_CNTL].gpu_addr);
1229 data = REG_SET_FIELD(0, WD_CNTL_SB_BUF_BASE, BASE, base);
1230 WREG32_SOC15(GC, 0, mmWD_CNTL_SB_BUF_BASE, data);
1232 base = upper_32_bits(adev->gfx.ngg.buf[NGG_CNTL].gpu_addr);
1233 data = REG_SET_FIELD(0, WD_CNTL_SB_BUF_BASE_HI, BASE_HI, base);
1234 WREG32_SOC15(GC, 0, mmWD_CNTL_SB_BUF_BASE_HI, data);
1236 /* Clear GDS reserved memory */
1237 r = amdgpu_ring_alloc(ring, 17);
1239 DRM_ERROR("amdgpu: NGG failed to lock ring %d (%d).\n",
1244 gfx_v9_0_write_data_to_reg(ring, 0, false,
1245 amdgpu_gds_reg_offset[0].mem_size,
1246 (adev->gds.mem.total_size +
1247 adev->gfx.ngg.gds_reserve_size) >>
1250 amdgpu_ring_write(ring, PACKET3(PACKET3_DMA_DATA, 5));
1251 amdgpu_ring_write(ring, (PACKET3_DMA_DATA_CP_SYNC |
1252 PACKET3_DMA_DATA_SRC_SEL(2)));
1253 amdgpu_ring_write(ring, 0);
1254 amdgpu_ring_write(ring, 0);
1255 amdgpu_ring_write(ring, adev->gfx.ngg.gds_reserve_addr);
1256 amdgpu_ring_write(ring, 0);
1257 amdgpu_ring_write(ring, adev->gfx.ngg.gds_reserve_size);
1260 gfx_v9_0_write_data_to_reg(ring, 0, false,
1261 amdgpu_gds_reg_offset[0].mem_size, 0);
1263 amdgpu_ring_commit(ring);
1268 static int gfx_v9_0_compute_ring_init(struct amdgpu_device *adev, int ring_id,
1269 int mec, int pipe, int queue)
1273 struct amdgpu_ring *ring = &adev->gfx.compute_ring[ring_id];
1275 ring = &adev->gfx.compute_ring[ring_id];
1280 ring->queue = queue;
1282 ring->ring_obj = NULL;
1283 ring->use_doorbell = true;
1284 ring->doorbell_index = (AMDGPU_DOORBELL_MEC_RING0 + ring_id) << 1;
1285 ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr
1286 + (ring_id * GFX9_MEC_HPD_SIZE);
1287 sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue);
1289 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP
1290 + ((ring->me - 1) * adev->gfx.mec.num_pipe_per_mec)
1293 /* type-2 packets are deprecated on MEC, use type-3 instead */
1294 r = amdgpu_ring_init(adev, ring, 1024,
1295 &adev->gfx.eop_irq, irq_type);
1303 static int gfx_v9_0_sw_init(void *handle)
1305 int i, j, k, r, ring_id;
1306 struct amdgpu_ring *ring;
1307 struct amdgpu_kiq *kiq;
1308 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1310 switch (adev->asic_type) {
1313 adev->gfx.mec.num_mec = 2;
1316 adev->gfx.mec.num_mec = 1;
1320 adev->gfx.mec.num_pipe_per_mec = 4;
1321 adev->gfx.mec.num_queue_per_pipe = 8;
1324 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_GRBM_CP, 178, &adev->gfx.kiq.irq);
1329 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_GRBM_CP, 181, &adev->gfx.eop_irq);
1333 /* Privileged reg */
1334 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_GRBM_CP, 184,
1335 &adev->gfx.priv_reg_irq);
1339 /* Privileged inst */
1340 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_GRBM_CP, 185,
1341 &adev->gfx.priv_inst_irq);
1345 adev->gfx.gfx_current_status = AMDGPU_GFX_NORMAL_MODE;
1347 gfx_v9_0_scratch_init(adev);
1349 r = gfx_v9_0_init_microcode(adev);
1351 DRM_ERROR("Failed to load gfx firmware!\n");
1355 r = gfx_v9_0_rlc_init(adev);
1357 DRM_ERROR("Failed to init rlc BOs!\n");
1361 r = gfx_v9_0_mec_init(adev);
1363 DRM_ERROR("Failed to init MEC BOs!\n");
1367 /* set up the gfx ring */
1368 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
1369 ring = &adev->gfx.gfx_ring[i];
1370 ring->ring_obj = NULL;
1371 sprintf(ring->name, "gfx");
1372 ring->use_doorbell = true;
1373 ring->doorbell_index = AMDGPU_DOORBELL64_GFX_RING0 << 1;
1374 r = amdgpu_ring_init(adev, ring, 1024,
1375 &adev->gfx.eop_irq, AMDGPU_CP_IRQ_GFX_EOP);
1380 /* set up the compute queues - allocate horizontally across pipes */
1382 for (i = 0; i < adev->gfx.mec.num_mec; ++i) {
1383 for (j = 0; j < adev->gfx.mec.num_queue_per_pipe; j++) {
1384 for (k = 0; k < adev->gfx.mec.num_pipe_per_mec; k++) {
1385 if (!amdgpu_gfx_is_mec_queue_enabled(adev, i, k, j))
1388 r = gfx_v9_0_compute_ring_init(adev,
1399 r = amdgpu_gfx_kiq_init(adev, GFX9_MEC_HPD_SIZE);
1401 DRM_ERROR("Failed to init KIQ BOs!\n");
1405 kiq = &adev->gfx.kiq;
1406 r = amdgpu_gfx_kiq_init_ring(adev, &kiq->ring, &kiq->irq);
1410 /* create MQD for all compute queues as wel as KIQ for SRIOV case */
1411 r = amdgpu_gfx_compute_mqd_sw_init(adev, sizeof(struct v9_mqd));
1415 /* reserve GDS, GWS and OA resource for gfx */
1416 r = amdgpu_bo_create_kernel(adev, adev->gds.mem.gfx_partition_size,
1417 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GDS,
1418 &adev->gds.gds_gfx_bo, NULL, NULL);
1422 r = amdgpu_bo_create_kernel(adev, adev->gds.gws.gfx_partition_size,
1423 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GWS,
1424 &adev->gds.gws_gfx_bo, NULL, NULL);
1428 r = amdgpu_bo_create_kernel(adev, adev->gds.oa.gfx_partition_size,
1429 PAGE_SIZE, AMDGPU_GEM_DOMAIN_OA,
1430 &adev->gds.oa_gfx_bo, NULL, NULL);
1434 adev->gfx.ce_ram_size = 0x8000;
1436 gfx_v9_0_gpu_early_init(adev);
1438 r = gfx_v9_0_ngg_init(adev);
1446 static int gfx_v9_0_sw_fini(void *handle)
1449 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1451 amdgpu_bo_free_kernel(&adev->gds.oa_gfx_bo, NULL, NULL);
1452 amdgpu_bo_free_kernel(&adev->gds.gws_gfx_bo, NULL, NULL);
1453 amdgpu_bo_free_kernel(&adev->gds.gds_gfx_bo, NULL, NULL);
1455 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
1456 amdgpu_ring_fini(&adev->gfx.gfx_ring[i]);
1457 for (i = 0; i < adev->gfx.num_compute_rings; i++)
1458 amdgpu_ring_fini(&adev->gfx.compute_ring[i]);
1460 amdgpu_gfx_compute_mqd_sw_fini(adev);
1461 amdgpu_gfx_kiq_free_ring(&adev->gfx.kiq.ring, &adev->gfx.kiq.irq);
1462 amdgpu_gfx_kiq_fini(adev);
1464 gfx_v9_0_mec_fini(adev);
1465 gfx_v9_0_ngg_fini(adev);
1471 static void gfx_v9_0_tiling_mode_table_init(struct amdgpu_device *adev)
1476 static void gfx_v9_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, u32 sh_num, u32 instance)
1480 if (instance == 0xffffffff)
1481 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_BROADCAST_WRITES, 1);
1483 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_INDEX, instance);
1485 if (se_num == 0xffffffff)
1486 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_BROADCAST_WRITES, 1);
1488 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_INDEX, se_num);
1490 if (sh_num == 0xffffffff)
1491 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SH_BROADCAST_WRITES, 1);
1493 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SH_INDEX, sh_num);
1495 WREG32_SOC15(GC, 0, mmGRBM_GFX_INDEX, data);
1498 static u32 gfx_v9_0_get_rb_active_bitmap(struct amdgpu_device *adev)
1502 data = RREG32_SOC15(GC, 0, mmCC_RB_BACKEND_DISABLE);
1503 data |= RREG32_SOC15(GC, 0, mmGC_USER_RB_BACKEND_DISABLE);
1505 data &= CC_RB_BACKEND_DISABLE__BACKEND_DISABLE_MASK;
1506 data >>= GC_USER_RB_BACKEND_DISABLE__BACKEND_DISABLE__SHIFT;
1508 mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_backends_per_se /
1509 adev->gfx.config.max_sh_per_se);
1511 return (~data) & mask;
1514 static void gfx_v9_0_setup_rb(struct amdgpu_device *adev)
1519 u32 rb_bitmap_width_per_sh = adev->gfx.config.max_backends_per_se /
1520 adev->gfx.config.max_sh_per_se;
1522 mutex_lock(&adev->grbm_idx_mutex);
1523 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
1524 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
1525 gfx_v9_0_select_se_sh(adev, i, j, 0xffffffff);
1526 data = gfx_v9_0_get_rb_active_bitmap(adev);
1527 active_rbs |= data << ((i * adev->gfx.config.max_sh_per_se + j) *
1528 rb_bitmap_width_per_sh);
1531 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
1532 mutex_unlock(&adev->grbm_idx_mutex);
1534 adev->gfx.config.backend_enable_mask = active_rbs;
1535 adev->gfx.config.num_rbs = hweight32(active_rbs);
1538 #define DEFAULT_SH_MEM_BASES (0x6000)
1539 #define FIRST_COMPUTE_VMID (8)
1540 #define LAST_COMPUTE_VMID (16)
1541 static void gfx_v9_0_init_compute_vmid(struct amdgpu_device *adev)
1544 uint32_t sh_mem_config;
1545 uint32_t sh_mem_bases;
1548 * Configure apertures:
1549 * LDS: 0x60000000'00000000 - 0x60000001'00000000 (4GB)
1550 * Scratch: 0x60000001'00000000 - 0x60000002'00000000 (4GB)
1551 * GPUVM: 0x60010000'00000000 - 0x60020000'00000000 (1TB)
1553 sh_mem_bases = DEFAULT_SH_MEM_BASES | (DEFAULT_SH_MEM_BASES << 16);
1555 sh_mem_config = SH_MEM_ADDRESS_MODE_64 |
1556 SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
1557 SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT;
1559 mutex_lock(&adev->srbm_mutex);
1560 for (i = FIRST_COMPUTE_VMID; i < LAST_COMPUTE_VMID; i++) {
1561 soc15_grbm_select(adev, 0, 0, 0, i);
1562 /* CP and shaders */
1563 WREG32_SOC15(GC, 0, mmSH_MEM_CONFIG, sh_mem_config);
1564 WREG32_SOC15(GC, 0, mmSH_MEM_BASES, sh_mem_bases);
1566 soc15_grbm_select(adev, 0, 0, 0, 0);
1567 mutex_unlock(&adev->srbm_mutex);
1570 static void gfx_v9_0_gpu_init(struct amdgpu_device *adev)
1575 WREG32_FIELD15(GC, 0, GRBM_CNTL, READ_TIMEOUT, 0xff);
1577 gfx_v9_0_tiling_mode_table_init(adev);
1579 gfx_v9_0_setup_rb(adev);
1580 gfx_v9_0_get_cu_info(adev, &adev->gfx.cu_info);
1582 /* XXX SH_MEM regs */
1583 /* where to put LDS, scratch, GPUVM in FSA64 space */
1584 mutex_lock(&adev->srbm_mutex);
1585 for (i = 0; i < 16; i++) {
1586 soc15_grbm_select(adev, 0, 0, 0, i);
1587 /* CP and shaders */
1589 tmp = REG_SET_FIELD(tmp, SH_MEM_CONFIG, ALIGNMENT_MODE,
1590 SH_MEM_ALIGNMENT_MODE_UNALIGNED);
1591 WREG32_SOC15(GC, 0, mmSH_MEM_CONFIG, tmp);
1592 WREG32_SOC15(GC, 0, mmSH_MEM_BASES, 0);
1594 soc15_grbm_select(adev, 0, 0, 0, 0);
1596 mutex_unlock(&adev->srbm_mutex);
1598 gfx_v9_0_init_compute_vmid(adev);
1600 mutex_lock(&adev->grbm_idx_mutex);
1602 * making sure that the following register writes will be broadcasted
1603 * to all the shaders
1605 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
1607 WREG32_SOC15(GC, 0, mmPA_SC_FIFO_SIZE,
1608 (adev->gfx.config.sc_prim_fifo_size_frontend <<
1609 PA_SC_FIFO_SIZE__SC_FRONTEND_PRIM_FIFO_SIZE__SHIFT) |
1610 (adev->gfx.config.sc_prim_fifo_size_backend <<
1611 PA_SC_FIFO_SIZE__SC_BACKEND_PRIM_FIFO_SIZE__SHIFT) |
1612 (adev->gfx.config.sc_hiz_tile_fifo_size <<
1613 PA_SC_FIFO_SIZE__SC_HIZ_TILE_FIFO_SIZE__SHIFT) |
1614 (adev->gfx.config.sc_earlyz_tile_fifo_size <<
1615 PA_SC_FIFO_SIZE__SC_EARLYZ_TILE_FIFO_SIZE__SHIFT));
1616 mutex_unlock(&adev->grbm_idx_mutex);
1620 static void gfx_v9_0_wait_for_rlc_serdes(struct amdgpu_device *adev)
1625 mutex_lock(&adev->grbm_idx_mutex);
1626 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
1627 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
1628 gfx_v9_0_select_se_sh(adev, i, j, 0xffffffff);
1629 for (k = 0; k < adev->usec_timeout; k++) {
1630 if (RREG32_SOC15(GC, 0, mmRLC_SERDES_CU_MASTER_BUSY) == 0)
1636 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
1637 mutex_unlock(&adev->grbm_idx_mutex);
1639 mask = RLC_SERDES_NONCU_MASTER_BUSY__SE_MASTER_BUSY_MASK |
1640 RLC_SERDES_NONCU_MASTER_BUSY__GC_MASTER_BUSY_MASK |
1641 RLC_SERDES_NONCU_MASTER_BUSY__TC0_MASTER_BUSY_MASK |
1642 RLC_SERDES_NONCU_MASTER_BUSY__TC1_MASTER_BUSY_MASK;
1643 for (k = 0; k < adev->usec_timeout; k++) {
1644 if ((RREG32_SOC15(GC, 0, mmRLC_SERDES_NONCU_MASTER_BUSY) & mask) == 0)
1650 static void gfx_v9_0_enable_gui_idle_interrupt(struct amdgpu_device *adev,
1653 u32 tmp = RREG32_SOC15(GC, 0, mmCP_INT_CNTL_RING0);
1655 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_BUSY_INT_ENABLE, enable ? 1 : 0);
1656 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_EMPTY_INT_ENABLE, enable ? 1 : 0);
1657 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CMP_BUSY_INT_ENABLE, enable ? 1 : 0);
1658 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, GFX_IDLE_INT_ENABLE, enable ? 1 : 0);
1660 WREG32_SOC15(GC, 0, mmCP_INT_CNTL_RING0, tmp);
1663 static void gfx_v9_0_init_csb(struct amdgpu_device *adev)
1666 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CSIB_ADDR_HI),
1667 adev->gfx.rlc.clear_state_gpu_addr >> 32);
1668 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CSIB_ADDR_LO),
1669 adev->gfx.rlc.clear_state_gpu_addr & 0xfffffffc);
1670 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_CSIB_LENGTH),
1671 adev->gfx.rlc.clear_state_size);
1674 static void gfx_v9_0_parse_ind_reg_list(int *register_list_format,
1675 int indirect_offset,
1677 int *unique_indirect_regs,
1678 int *unique_indirect_reg_count,
1679 int max_indirect_reg_count,
1680 int *indirect_start_offsets,
1681 int *indirect_start_offsets_count,
1682 int max_indirect_start_offsets_count)
1685 bool new_entry = true;
1687 for (; indirect_offset < list_size; indirect_offset++) {
1691 indirect_start_offsets[*indirect_start_offsets_count] = indirect_offset;
1692 *indirect_start_offsets_count = *indirect_start_offsets_count + 1;
1693 BUG_ON(*indirect_start_offsets_count >= max_indirect_start_offsets_count);
1696 if (register_list_format[indirect_offset] == 0xFFFFFFFF) {
1701 indirect_offset += 2;
1703 /* look for the matching indice */
1704 for (idx = 0; idx < *unique_indirect_reg_count; idx++) {
1705 if (unique_indirect_regs[idx] ==
1706 register_list_format[indirect_offset])
1710 if (idx >= *unique_indirect_reg_count) {
1711 unique_indirect_regs[*unique_indirect_reg_count] =
1712 register_list_format[indirect_offset];
1713 idx = *unique_indirect_reg_count;
1714 *unique_indirect_reg_count = *unique_indirect_reg_count + 1;
1715 BUG_ON(*unique_indirect_reg_count >= max_indirect_reg_count);
1718 register_list_format[indirect_offset] = idx;
1722 static int gfx_v9_0_init_rlc_save_restore_list(struct amdgpu_device *adev)
1724 int unique_indirect_regs[] = {0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0};
1725 int unique_indirect_reg_count = 0;
1727 int indirect_start_offsets[] = {0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0};
1728 int indirect_start_offsets_count = 0;
1734 u32 *register_list_format =
1735 kmalloc(adev->gfx.rlc.reg_list_format_size_bytes, GFP_KERNEL);
1736 if (!register_list_format)
1738 memcpy(register_list_format, adev->gfx.rlc.register_list_format,
1739 adev->gfx.rlc.reg_list_format_size_bytes);
1741 /* setup unique_indirect_regs array and indirect_start_offsets array */
1742 gfx_v9_0_parse_ind_reg_list(register_list_format,
1743 GFX9_RLC_FORMAT_DIRECT_REG_LIST_LENGTH,
1744 adev->gfx.rlc.reg_list_format_size_bytes >> 2,
1745 unique_indirect_regs,
1746 &unique_indirect_reg_count,
1747 sizeof(unique_indirect_regs)/sizeof(int),
1748 indirect_start_offsets,
1749 &indirect_start_offsets_count,
1750 sizeof(indirect_start_offsets)/sizeof(int));
1752 /* enable auto inc in case it is disabled */
1753 tmp = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_CNTL));
1754 tmp |= RLC_SRM_CNTL__AUTO_INCR_ADDR_MASK;
1755 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_CNTL), tmp);
1757 /* write register_restore table to offset 0x0 using RLC_SRM_ARAM_ADDR/DATA */
1758 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_ARAM_ADDR),
1759 RLC_SAVE_RESTORE_ADDR_STARTING_OFFSET);
1760 for (i = 0; i < adev->gfx.rlc.reg_list_size_bytes >> 2; i++)
1761 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_ARAM_DATA),
1762 adev->gfx.rlc.register_restore[i]);
1764 /* load direct register */
1765 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_ARAM_ADDR), 0);
1766 for (i = 0; i < adev->gfx.rlc.reg_list_size_bytes >> 2; i++)
1767 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_ARAM_DATA),
1768 adev->gfx.rlc.register_restore[i]);
1770 /* load indirect register */
1771 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_SCRATCH_ADDR),
1772 adev->gfx.rlc.reg_list_format_start);
1773 for (i = 0; i < adev->gfx.rlc.reg_list_format_size_bytes >> 2; i++)
1774 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_SCRATCH_DATA),
1775 register_list_format[i]);
1777 /* set save/restore list size */
1778 list_size = adev->gfx.rlc.reg_list_size_bytes >> 2;
1779 list_size = list_size >> 1;
1780 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_SCRATCH_ADDR),
1781 adev->gfx.rlc.reg_restore_list_size);
1782 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_SCRATCH_DATA), list_size);
1784 /* write the starting offsets to RLC scratch ram */
1785 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_SCRATCH_ADDR),
1786 adev->gfx.rlc.starting_offsets_start);
1787 for (i = 0; i < sizeof(indirect_start_offsets)/sizeof(int); i++)
1788 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_GPM_SCRATCH_DATA),
1789 indirect_start_offsets[i]);
1791 /* load unique indirect regs*/
1792 for (i = 0; i < sizeof(unique_indirect_regs)/sizeof(int); i++) {
1793 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_INDEX_CNTL_ADDR_0) + i,
1794 unique_indirect_regs[i] & 0x3FFFF);
1795 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_INDEX_CNTL_DATA_0) + i,
1796 unique_indirect_regs[i] >> 20);
1799 kfree(register_list_format);
1803 static void gfx_v9_0_enable_save_restore_machine(struct amdgpu_device *adev)
1807 tmp = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_CNTL));
1808 tmp |= RLC_SRM_CNTL__SRM_ENABLE_MASK;
1809 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_SRM_CNTL), tmp);
1812 static void pwr_10_0_gfxip_control_over_cgpg(struct amdgpu_device *adev,
1816 uint32_t default_data = 0;
1818 default_data = data = RREG32(SOC15_REG_OFFSET(PWR, 0, mmPWR_MISC_CNTL_STATUS));
1819 if (enable == true) {
1820 /* enable GFXIP control over CGPG */
1821 data |= PWR_MISC_CNTL_STATUS__PWR_GFX_RLC_CGPG_EN_MASK;
1822 if(default_data != data)
1823 WREG32(SOC15_REG_OFFSET(PWR, 0, mmPWR_MISC_CNTL_STATUS), data);
1826 data &= ~PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS_MASK;
1827 data |= (2 << PWR_MISC_CNTL_STATUS__PWR_GFXOFF_STATUS__SHIFT);
1828 if(default_data != data)
1829 WREG32(SOC15_REG_OFFSET(PWR, 0, mmPWR_MISC_CNTL_STATUS), data);
1831 /* restore GFXIP control over GCPG */
1832 data &= ~PWR_MISC_CNTL_STATUS__PWR_GFX_RLC_CGPG_EN_MASK;
1833 if(default_data != data)
1834 WREG32(SOC15_REG_OFFSET(PWR, 0, mmPWR_MISC_CNTL_STATUS), data);
1838 static void gfx_v9_0_init_gfx_power_gating(struct amdgpu_device *adev)
1842 if (adev->pg_flags & (AMD_PG_SUPPORT_GFX_PG |
1843 AMD_PG_SUPPORT_GFX_SMG |
1844 AMD_PG_SUPPORT_GFX_DMG)) {
1845 /* init IDLE_POLL_COUNT = 60 */
1846 data = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_RB_WPTR_POLL_CNTL));
1847 data &= ~CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT_MASK;
1848 data |= (0x60 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT);
1849 WREG32(SOC15_REG_OFFSET(GC, 0, mmCP_RB_WPTR_POLL_CNTL), data);
1851 /* init RLC PG Delay */
1853 data |= (0x10 << RLC_PG_DELAY__POWER_UP_DELAY__SHIFT);
1854 data |= (0x10 << RLC_PG_DELAY__POWER_DOWN_DELAY__SHIFT);
1855 data |= (0x10 << RLC_PG_DELAY__CMD_PROPAGATE_DELAY__SHIFT);
1856 data |= (0x40 << RLC_PG_DELAY__MEM_SLEEP_DELAY__SHIFT);
1857 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_DELAY), data);
1859 data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_DELAY_2));
1860 data &= ~RLC_PG_DELAY_2__SERDES_CMD_DELAY_MASK;
1861 data |= (0x4 << RLC_PG_DELAY_2__SERDES_CMD_DELAY__SHIFT);
1862 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_DELAY_2), data);
1864 data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_DELAY_3));
1865 data &= ~RLC_PG_DELAY_3__CGCG_ACTIVE_BEFORE_CGPG_MASK;
1866 data |= (0xff << RLC_PG_DELAY_3__CGCG_ACTIVE_BEFORE_CGPG__SHIFT);
1867 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_DELAY_3), data);
1869 data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_AUTO_PG_CTRL));
1870 data &= ~RLC_AUTO_PG_CTRL__GRBM_REG_SAVE_GFX_IDLE_THRESHOLD_MASK;
1872 /* program GRBM_REG_SAVE_GFX_IDLE_THRESHOLD to 0x55f0 */
1873 data |= (0x55f0 << RLC_AUTO_PG_CTRL__GRBM_REG_SAVE_GFX_IDLE_THRESHOLD__SHIFT);
1874 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_AUTO_PG_CTRL), data);
1876 pwr_10_0_gfxip_control_over_cgpg(adev, true);
1880 static void gfx_v9_0_enable_sck_slow_down_on_power_up(struct amdgpu_device *adev,
1884 uint32_t default_data = 0;
1886 default_data = data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL));
1888 if (enable == true) {
1889 data |= RLC_PG_CNTL__SMU_CLK_SLOWDOWN_ON_PU_ENABLE_MASK;
1890 if (default_data != data)
1891 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1893 data &= ~RLC_PG_CNTL__SMU_CLK_SLOWDOWN_ON_PU_ENABLE_MASK;
1894 if(default_data != data)
1895 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1899 static void gfx_v9_0_enable_sck_slow_down_on_power_down(struct amdgpu_device *adev,
1903 uint32_t default_data = 0;
1905 default_data = data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL));
1907 if (enable == true) {
1908 data |= RLC_PG_CNTL__SMU_CLK_SLOWDOWN_ON_PD_ENABLE_MASK;
1909 if(default_data != data)
1910 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1912 data &= ~RLC_PG_CNTL__SMU_CLK_SLOWDOWN_ON_PD_ENABLE_MASK;
1913 if(default_data != data)
1914 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1918 static void gfx_v9_0_enable_cp_power_gating(struct amdgpu_device *adev,
1922 uint32_t default_data = 0;
1924 default_data = data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL));
1926 if (enable == true) {
1927 data &= ~RLC_PG_CNTL__CP_PG_DISABLE_MASK;
1928 if(default_data != data)
1929 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1931 data |= RLC_PG_CNTL__CP_PG_DISABLE_MASK;
1932 if(default_data != data)
1933 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1937 static void gfx_v9_0_enable_gfx_cg_power_gating(struct amdgpu_device *adev,
1940 uint32_t data, default_data;
1942 default_data = data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL));
1944 data |= RLC_PG_CNTL__GFX_POWER_GATING_ENABLE_MASK;
1946 data &= ~RLC_PG_CNTL__GFX_POWER_GATING_ENABLE_MASK;
1947 if(default_data != data)
1948 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1951 static void gfx_v9_0_enable_gfx_pipeline_powergating(struct amdgpu_device *adev,
1954 uint32_t data, default_data;
1956 default_data = data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL));
1958 data |= RLC_PG_CNTL__GFX_PIPELINE_PG_ENABLE_MASK;
1960 data &= ~RLC_PG_CNTL__GFX_PIPELINE_PG_ENABLE_MASK;
1961 if(default_data != data)
1962 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1965 /* read any GFX register to wake up GFX */
1966 data = RREG32(SOC15_REG_OFFSET(GC, 0, mmDB_RENDER_CONTROL));
1969 static void gfx_v9_0_enable_gfx_static_mg_power_gating(struct amdgpu_device *adev,
1972 uint32_t data, default_data;
1974 default_data = data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL));
1976 data |= RLC_PG_CNTL__STATIC_PER_CU_PG_ENABLE_MASK;
1978 data &= ~RLC_PG_CNTL__STATIC_PER_CU_PG_ENABLE_MASK;
1979 if(default_data != data)
1980 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1983 static void gfx_v9_0_enable_gfx_dynamic_mg_power_gating(struct amdgpu_device *adev,
1986 uint32_t data, default_data;
1988 default_data = data = RREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL));
1990 data |= RLC_PG_CNTL__DYN_PER_CU_PG_ENABLE_MASK;
1992 data &= ~RLC_PG_CNTL__DYN_PER_CU_PG_ENABLE_MASK;
1993 if(default_data != data)
1994 WREG32(SOC15_REG_OFFSET(GC, 0, mmRLC_PG_CNTL), data);
1997 static void gfx_v9_0_init_pg(struct amdgpu_device *adev)
1999 if (adev->pg_flags & (AMD_PG_SUPPORT_GFX_PG |
2000 AMD_PG_SUPPORT_GFX_SMG |
2001 AMD_PG_SUPPORT_GFX_DMG |
2003 AMD_PG_SUPPORT_GDS |
2004 AMD_PG_SUPPORT_RLC_SMU_HS)) {
2005 gfx_v9_0_init_csb(adev);
2006 gfx_v9_0_init_rlc_save_restore_list(adev);
2007 gfx_v9_0_enable_save_restore_machine(adev);
2009 if (adev->asic_type == CHIP_RAVEN) {
2010 WREG32(mmRLC_JUMP_TABLE_RESTORE,
2011 adev->gfx.rlc.cp_table_gpu_addr >> 8);
2012 gfx_v9_0_init_gfx_power_gating(adev);
2014 if (adev->pg_flags & AMD_PG_SUPPORT_RLC_SMU_HS) {
2015 gfx_v9_0_enable_sck_slow_down_on_power_up(adev, true);
2016 gfx_v9_0_enable_sck_slow_down_on_power_down(adev, true);
2018 gfx_v9_0_enable_sck_slow_down_on_power_up(adev, false);
2019 gfx_v9_0_enable_sck_slow_down_on_power_down(adev, false);
2022 if (adev->pg_flags & AMD_PG_SUPPORT_CP)
2023 gfx_v9_0_enable_cp_power_gating(adev, true);
2025 gfx_v9_0_enable_cp_power_gating(adev, false);
2030 void gfx_v9_0_rlc_stop(struct amdgpu_device *adev)
2032 u32 tmp = RREG32_SOC15(GC, 0, mmRLC_CNTL);
2034 tmp = REG_SET_FIELD(tmp, RLC_CNTL, RLC_ENABLE_F32, 0);
2035 WREG32_SOC15(GC, 0, mmRLC_CNTL, tmp);
2037 gfx_v9_0_enable_gui_idle_interrupt(adev, false);
2039 gfx_v9_0_wait_for_rlc_serdes(adev);
2042 static void gfx_v9_0_rlc_reset(struct amdgpu_device *adev)
2044 WREG32_FIELD15(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 1);
2046 WREG32_FIELD15(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 0);
2050 static void gfx_v9_0_rlc_start(struct amdgpu_device *adev)
2052 #ifdef AMDGPU_RLC_DEBUG_RETRY
2056 WREG32_FIELD15(GC, 0, RLC_CNTL, RLC_ENABLE_F32, 1);
2058 /* carrizo do enable cp interrupt after cp inited */
2059 if (!(adev->flags & AMD_IS_APU))
2060 gfx_v9_0_enable_gui_idle_interrupt(adev, true);
2064 #ifdef AMDGPU_RLC_DEBUG_RETRY
2065 /* RLC_GPM_GENERAL_6 : RLC Ucode version */
2066 rlc_ucode_ver = RREG32_SOC15(GC, 0, mmRLC_GPM_GENERAL_6);
2067 if(rlc_ucode_ver == 0x108) {
2068 DRM_INFO("Using rlc debug ucode. mmRLC_GPM_GENERAL_6 ==0x08%x / fw_ver == %i \n",
2069 rlc_ucode_ver, adev->gfx.rlc_fw_version);
2070 /* RLC_GPM_TIMER_INT_3 : Timer interval in RefCLK cycles,
2071 * default is 0x9C4 to create a 100us interval */
2072 WREG32_SOC15(GC, 0, mmRLC_GPM_TIMER_INT_3, 0x9C4);
2073 /* RLC_GPM_GENERAL_12 : Minimum gap between wptr and rptr
2074 * to disable the page fault retry interrupts, default is
2076 WREG32_SOC15(GC, 0, mmRLC_GPM_GENERAL_12, 0x100);
2081 static int gfx_v9_0_rlc_load_microcode(struct amdgpu_device *adev)
2083 const struct rlc_firmware_header_v2_0 *hdr;
2084 const __le32 *fw_data;
2085 unsigned i, fw_size;
2087 if (!adev->gfx.rlc_fw)
2090 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data;
2091 amdgpu_ucode_print_rlc_hdr(&hdr->header);
2093 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
2094 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2095 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
2097 WREG32_SOC15(GC, 0, mmRLC_GPM_UCODE_ADDR,
2098 RLCG_UCODE_LOADING_START_ADDRESS);
2099 for (i = 0; i < fw_size; i++)
2100 WREG32_SOC15(GC, 0, mmRLC_GPM_UCODE_DATA, le32_to_cpup(fw_data++));
2101 WREG32_SOC15(GC, 0, mmRLC_GPM_UCODE_ADDR, adev->gfx.rlc_fw_version);
2106 static int gfx_v9_0_rlc_resume(struct amdgpu_device *adev)
2110 if (amdgpu_sriov_vf(adev))
2113 gfx_v9_0_rlc_stop(adev);
2116 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL, 0);
2119 WREG32_SOC15(GC, 0, mmRLC_PG_CNTL, 0);
2121 gfx_v9_0_rlc_reset(adev);
2123 gfx_v9_0_init_pg(adev);
2125 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
2126 /* legacy rlc firmware loading */
2127 r = gfx_v9_0_rlc_load_microcode(adev);
2132 if (adev->asic_type == CHIP_RAVEN) {
2133 if (amdgpu_lbpw != 0)
2134 gfx_v9_0_enable_lbpw(adev, true);
2136 gfx_v9_0_enable_lbpw(adev, false);
2139 gfx_v9_0_rlc_start(adev);
2144 static void gfx_v9_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable)
2147 u32 tmp = RREG32_SOC15(GC, 0, mmCP_ME_CNTL);
2149 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_HALT, enable ? 0 : 1);
2150 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_HALT, enable ? 0 : 1);
2151 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, CE_HALT, enable ? 0 : 1);
2153 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
2154 adev->gfx.gfx_ring[i].ready = false;
2156 WREG32_SOC15(GC, 0, mmCP_ME_CNTL, tmp);
2160 static int gfx_v9_0_cp_gfx_load_microcode(struct amdgpu_device *adev)
2162 const struct gfx_firmware_header_v1_0 *pfp_hdr;
2163 const struct gfx_firmware_header_v1_0 *ce_hdr;
2164 const struct gfx_firmware_header_v1_0 *me_hdr;
2165 const __le32 *fw_data;
2166 unsigned i, fw_size;
2168 if (!adev->gfx.me_fw || !adev->gfx.pfp_fw || !adev->gfx.ce_fw)
2171 pfp_hdr = (const struct gfx_firmware_header_v1_0 *)
2172 adev->gfx.pfp_fw->data;
2173 ce_hdr = (const struct gfx_firmware_header_v1_0 *)
2174 adev->gfx.ce_fw->data;
2175 me_hdr = (const struct gfx_firmware_header_v1_0 *)
2176 adev->gfx.me_fw->data;
2178 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header);
2179 amdgpu_ucode_print_gfx_hdr(&ce_hdr->header);
2180 amdgpu_ucode_print_gfx_hdr(&me_hdr->header);
2182 gfx_v9_0_cp_gfx_enable(adev, false);
2185 fw_data = (const __le32 *)
2186 (adev->gfx.pfp_fw->data +
2187 le32_to_cpu(pfp_hdr->header.ucode_array_offset_bytes));
2188 fw_size = le32_to_cpu(pfp_hdr->header.ucode_size_bytes) / 4;
2189 WREG32_SOC15(GC, 0, mmCP_PFP_UCODE_ADDR, 0);
2190 for (i = 0; i < fw_size; i++)
2191 WREG32_SOC15(GC, 0, mmCP_PFP_UCODE_DATA, le32_to_cpup(fw_data++));
2192 WREG32_SOC15(GC, 0, mmCP_PFP_UCODE_ADDR, adev->gfx.pfp_fw_version);
2195 fw_data = (const __le32 *)
2196 (adev->gfx.ce_fw->data +
2197 le32_to_cpu(ce_hdr->header.ucode_array_offset_bytes));
2198 fw_size = le32_to_cpu(ce_hdr->header.ucode_size_bytes) / 4;
2199 WREG32_SOC15(GC, 0, mmCP_CE_UCODE_ADDR, 0);
2200 for (i = 0; i < fw_size; i++)
2201 WREG32_SOC15(GC, 0, mmCP_CE_UCODE_DATA, le32_to_cpup(fw_data++));
2202 WREG32_SOC15(GC, 0, mmCP_CE_UCODE_ADDR, adev->gfx.ce_fw_version);
2205 fw_data = (const __le32 *)
2206 (adev->gfx.me_fw->data +
2207 le32_to_cpu(me_hdr->header.ucode_array_offset_bytes));
2208 fw_size = le32_to_cpu(me_hdr->header.ucode_size_bytes) / 4;
2209 WREG32_SOC15(GC, 0, mmCP_ME_RAM_WADDR, 0);
2210 for (i = 0; i < fw_size; i++)
2211 WREG32_SOC15(GC, 0, mmCP_ME_RAM_DATA, le32_to_cpup(fw_data++));
2212 WREG32_SOC15(GC, 0, mmCP_ME_RAM_WADDR, adev->gfx.me_fw_version);
2217 static int gfx_v9_0_cp_gfx_start(struct amdgpu_device *adev)
2219 struct amdgpu_ring *ring = &adev->gfx.gfx_ring[0];
2220 const struct cs_section_def *sect = NULL;
2221 const struct cs_extent_def *ext = NULL;
2225 WREG32_SOC15(GC, 0, mmCP_MAX_CONTEXT, adev->gfx.config.max_hw_contexts - 1);
2226 WREG32_SOC15(GC, 0, mmCP_DEVICE_ID, 1);
2228 gfx_v9_0_cp_gfx_enable(adev, true);
2230 r = amdgpu_ring_alloc(ring, gfx_v9_0_get_csb_size(adev) + 4);
2232 DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r);
2236 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
2237 amdgpu_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
2239 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
2240 amdgpu_ring_write(ring, 0x80000000);
2241 amdgpu_ring_write(ring, 0x80000000);
2243 for (sect = gfx9_cs_data; sect->section != NULL; ++sect) {
2244 for (ext = sect->section; ext->extent != NULL; ++ext) {
2245 if (sect->id == SECT_CONTEXT) {
2246 amdgpu_ring_write(ring,
2247 PACKET3(PACKET3_SET_CONTEXT_REG,
2249 amdgpu_ring_write(ring,
2250 ext->reg_index - PACKET3_SET_CONTEXT_REG_START);
2251 for (i = 0; i < ext->reg_count; i++)
2252 amdgpu_ring_write(ring, ext->extent[i]);
2257 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
2258 amdgpu_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
2260 amdgpu_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
2261 amdgpu_ring_write(ring, 0);
2263 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
2264 amdgpu_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
2265 amdgpu_ring_write(ring, 0x8000);
2266 amdgpu_ring_write(ring, 0x8000);
2268 amdgpu_ring_commit(ring);
2273 static int gfx_v9_0_cp_gfx_resume(struct amdgpu_device *adev)
2275 struct amdgpu_ring *ring;
2278 u64 rb_addr, rptr_addr, wptr_gpu_addr;
2280 /* Set the write pointer delay */
2281 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_DELAY, 0);
2283 /* set the RB to use vmid 0 */
2284 WREG32_SOC15(GC, 0, mmCP_RB_VMID, 0);
2286 /* Set ring buffer size */
2287 ring = &adev->gfx.gfx_ring[0];
2288 rb_bufsz = order_base_2(ring->ring_size / 8);
2289 tmp = REG_SET_FIELD(0, CP_RB0_CNTL, RB_BUFSZ, rb_bufsz);
2290 tmp = REG_SET_FIELD(tmp, CP_RB0_CNTL, RB_BLKSZ, rb_bufsz - 2);
2292 tmp = REG_SET_FIELD(tmp, CP_RB0_CNTL, BUF_SWAP, 1);
2294 WREG32_SOC15(GC, 0, mmCP_RB0_CNTL, tmp);
2296 /* Initialize the ring buffer's write pointers */
2298 WREG32_SOC15(GC, 0, mmCP_RB0_WPTR, lower_32_bits(ring->wptr));
2299 WREG32_SOC15(GC, 0, mmCP_RB0_WPTR_HI, upper_32_bits(ring->wptr));
2301 /* set the wb address wether it's enabled or not */
2302 rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
2303 WREG32_SOC15(GC, 0, mmCP_RB0_RPTR_ADDR, lower_32_bits(rptr_addr));
2304 WREG32_SOC15(GC, 0, mmCP_RB0_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & CP_RB_RPTR_ADDR_HI__RB_RPTR_ADDR_HI_MASK);
2306 wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
2307 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_ADDR_LO, lower_32_bits(wptr_gpu_addr));
2308 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_ADDR_HI, upper_32_bits(wptr_gpu_addr));
2311 WREG32_SOC15(GC, 0, mmCP_RB0_CNTL, tmp);
2313 rb_addr = ring->gpu_addr >> 8;
2314 WREG32_SOC15(GC, 0, mmCP_RB0_BASE, rb_addr);
2315 WREG32_SOC15(GC, 0, mmCP_RB0_BASE_HI, upper_32_bits(rb_addr));
2317 tmp = RREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_CONTROL);
2318 if (ring->use_doorbell) {
2319 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
2320 DOORBELL_OFFSET, ring->doorbell_index);
2321 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
2324 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, DOORBELL_EN, 0);
2326 WREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_CONTROL, tmp);
2328 tmp = REG_SET_FIELD(0, CP_RB_DOORBELL_RANGE_LOWER,
2329 DOORBELL_RANGE_LOWER, ring->doorbell_index);
2330 WREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_RANGE_LOWER, tmp);
2332 WREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_RANGE_UPPER,
2333 CP_RB_DOORBELL_RANGE_UPPER__DOORBELL_RANGE_UPPER_MASK);
2336 /* start the ring */
2337 gfx_v9_0_cp_gfx_start(adev);
2343 static void gfx_v9_0_cp_compute_enable(struct amdgpu_device *adev, bool enable)
2348 WREG32_SOC15(GC, 0, mmCP_MEC_CNTL, 0);
2350 WREG32_SOC15(GC, 0, mmCP_MEC_CNTL,
2351 (CP_MEC_CNTL__MEC_ME1_HALT_MASK | CP_MEC_CNTL__MEC_ME2_HALT_MASK));
2352 for (i = 0; i < adev->gfx.num_compute_rings; i++)
2353 adev->gfx.compute_ring[i].ready = false;
2354 adev->gfx.kiq.ring.ready = false;
2359 static int gfx_v9_0_cp_compute_load_microcode(struct amdgpu_device *adev)
2361 const struct gfx_firmware_header_v1_0 *mec_hdr;
2362 const __le32 *fw_data;
2366 if (!adev->gfx.mec_fw)
2369 gfx_v9_0_cp_compute_enable(adev, false);
2371 mec_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
2372 amdgpu_ucode_print_gfx_hdr(&mec_hdr->header);
2374 fw_data = (const __le32 *)
2375 (adev->gfx.mec_fw->data +
2376 le32_to_cpu(mec_hdr->header.ucode_array_offset_bytes));
2378 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, VMID, 0);
2379 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, CACHE_POLICY, 0);
2380 WREG32_SOC15(GC, 0, mmCP_CPC_IC_BASE_CNTL, tmp);
2382 WREG32_SOC15(GC, 0, mmCP_CPC_IC_BASE_LO,
2383 adev->gfx.mec.mec_fw_gpu_addr & 0xFFFFF000);
2384 WREG32_SOC15(GC, 0, mmCP_CPC_IC_BASE_HI,
2385 upper_32_bits(adev->gfx.mec.mec_fw_gpu_addr));
2388 WREG32_SOC15(GC, 0, mmCP_MEC_ME1_UCODE_ADDR,
2389 mec_hdr->jt_offset);
2390 for (i = 0; i < mec_hdr->jt_size; i++)
2391 WREG32_SOC15(GC, 0, mmCP_MEC_ME1_UCODE_DATA,
2392 le32_to_cpup(fw_data + mec_hdr->jt_offset + i));
2394 WREG32_SOC15(GC, 0, mmCP_MEC_ME1_UCODE_ADDR,
2395 adev->gfx.mec_fw_version);
2396 /* Todo : Loading MEC2 firmware is only necessary if MEC2 should run different microcode than MEC1. */
2402 static void gfx_v9_0_kiq_setting(struct amdgpu_ring *ring)
2405 struct amdgpu_device *adev = ring->adev;
2407 /* tell RLC which is KIQ queue */
2408 tmp = RREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS);
2410 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
2411 WREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS, tmp);
2413 WREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS, tmp);
2416 static int gfx_v9_0_kiq_kcq_enable(struct amdgpu_device *adev)
2418 struct amdgpu_ring *kiq_ring = &adev->gfx.kiq.ring;
2419 uint32_t scratch, tmp = 0;
2420 uint64_t queue_mask = 0;
2423 for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
2424 if (!test_bit(i, adev->gfx.mec.queue_bitmap))
2427 /* This situation may be hit in the future if a new HW
2428 * generation exposes more than 64 queues. If so, the
2429 * definition of queue_mask needs updating */
2430 if (WARN_ON(i > (sizeof(queue_mask)*8))) {
2431 DRM_ERROR("Invalid KCQ enabled: %d\n", i);
2435 queue_mask |= (1ull << i);
2438 r = amdgpu_gfx_scratch_get(adev, &scratch);
2440 DRM_ERROR("Failed to get scratch reg (%d).\n", r);
2443 WREG32(scratch, 0xCAFEDEAD);
2445 r = amdgpu_ring_alloc(kiq_ring, (7 * adev->gfx.num_compute_rings) + 11);
2447 DRM_ERROR("Failed to lock KIQ (%d).\n", r);
2448 amdgpu_gfx_scratch_free(adev, scratch);
2453 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_SET_RESOURCES, 6));
2454 amdgpu_ring_write(kiq_ring, PACKET3_SET_RESOURCES_VMID_MASK(0) |
2455 PACKET3_SET_RESOURCES_QUEUE_TYPE(0)); /* vmid_mask:0 queue_type:0 (KIQ) */
2456 amdgpu_ring_write(kiq_ring, lower_32_bits(queue_mask)); /* queue mask lo */
2457 amdgpu_ring_write(kiq_ring, upper_32_bits(queue_mask)); /* queue mask hi */
2458 amdgpu_ring_write(kiq_ring, 0); /* gws mask lo */
2459 amdgpu_ring_write(kiq_ring, 0); /* gws mask hi */
2460 amdgpu_ring_write(kiq_ring, 0); /* oac mask */
2461 amdgpu_ring_write(kiq_ring, 0); /* gds heap base:0, gds heap size:0 */
2462 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
2463 struct amdgpu_ring *ring = &adev->gfx.compute_ring[i];
2464 uint64_t mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
2465 uint64_t wptr_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
2467 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
2468 /* Q_sel:0, vmid:0, vidmem: 1, engine:0, num_Q:1*/
2469 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */
2470 PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
2471 PACKET3_MAP_QUEUES_VMID(0) | /* VMID */
2472 PACKET3_MAP_QUEUES_QUEUE(ring->queue) |
2473 PACKET3_MAP_QUEUES_PIPE(ring->pipe) |
2474 PACKET3_MAP_QUEUES_ME((ring->me == 1 ? 0 : 1)) |
2475 PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
2476 PACKET3_MAP_QUEUES_ALLOC_FORMAT(1) | /* alloc format: all_on_one_pipe */
2477 PACKET3_MAP_QUEUES_ENGINE_SEL(0) | /* engine_sel: compute */
2478 PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
2479 amdgpu_ring_write(kiq_ring, PACKET3_MAP_QUEUES_DOORBELL_OFFSET(ring->doorbell_index));
2480 amdgpu_ring_write(kiq_ring, lower_32_bits(mqd_addr));
2481 amdgpu_ring_write(kiq_ring, upper_32_bits(mqd_addr));
2482 amdgpu_ring_write(kiq_ring, lower_32_bits(wptr_addr));
2483 amdgpu_ring_write(kiq_ring, upper_32_bits(wptr_addr));
2485 /* write to scratch for completion */
2486 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1));
2487 amdgpu_ring_write(kiq_ring, (scratch - PACKET3_SET_UCONFIG_REG_START));
2488 amdgpu_ring_write(kiq_ring, 0xDEADBEEF);
2489 amdgpu_ring_commit(kiq_ring);
2491 for (i = 0; i < adev->usec_timeout; i++) {
2492 tmp = RREG32(scratch);
2493 if (tmp == 0xDEADBEEF)
2497 if (i >= adev->usec_timeout) {
2498 DRM_ERROR("KCQ enable failed (scratch(0x%04X)=0x%08X)\n",
2502 amdgpu_gfx_scratch_free(adev, scratch);
2507 static int gfx_v9_0_mqd_init(struct amdgpu_ring *ring)
2509 struct amdgpu_device *adev = ring->adev;
2510 struct v9_mqd *mqd = ring->mqd_ptr;
2511 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
2514 mqd->header = 0xC0310800;
2515 mqd->compute_pipelinestat_enable = 0x00000001;
2516 mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
2517 mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
2518 mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
2519 mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
2520 mqd->compute_misc_reserved = 0x00000003;
2522 eop_base_addr = ring->eop_gpu_addr >> 8;
2523 mqd->cp_hqd_eop_base_addr_lo = eop_base_addr;
2524 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
2526 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
2527 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_EOP_CONTROL);
2528 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
2529 (order_base_2(GFX9_MEC_HPD_SIZE / 4) - 1));
2531 mqd->cp_hqd_eop_control = tmp;
2533 /* enable doorbell? */
2534 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL);
2536 if (ring->use_doorbell) {
2537 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2538 DOORBELL_OFFSET, ring->doorbell_index);
2539 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2541 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2542 DOORBELL_SOURCE, 0);
2543 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2547 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2550 mqd->cp_hqd_pq_doorbell_control = tmp;
2552 /* disable the queue if it's active */
2554 mqd->cp_hqd_dequeue_request = 0;
2555 mqd->cp_hqd_pq_rptr = 0;
2556 mqd->cp_hqd_pq_wptr_lo = 0;
2557 mqd->cp_hqd_pq_wptr_hi = 0;
2559 /* set the pointer to the MQD */
2560 mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
2561 mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
2563 /* set MQD vmid to 0 */
2564 tmp = RREG32_SOC15(GC, 0, mmCP_MQD_CONTROL);
2565 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
2566 mqd->cp_mqd_control = tmp;
2568 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
2569 hqd_gpu_addr = ring->gpu_addr >> 8;
2570 mqd->cp_hqd_pq_base_lo = hqd_gpu_addr;
2571 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
2573 /* set up the HQD, this is similar to CP_RB0_CNTL */
2574 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_CONTROL);
2575 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
2576 (order_base_2(ring->ring_size / 4) - 1));
2577 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
2578 ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
2580 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, ENDIAN_SWAP, 1);
2582 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 0);
2583 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, ROQ_PQ_IB_FLIP, 0);
2584 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
2585 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
2586 mqd->cp_hqd_pq_control = tmp;
2588 /* set the wb address whether it's enabled or not */
2589 wb_gpu_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
2590 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
2591 mqd->cp_hqd_pq_rptr_report_addr_hi =
2592 upper_32_bits(wb_gpu_addr) & 0xffff;
2594 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
2595 wb_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
2596 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc;
2597 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
2600 /* enable the doorbell if requested */
2601 if (ring->use_doorbell) {
2602 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL);
2603 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2604 DOORBELL_OFFSET, ring->doorbell_index);
2606 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2608 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2609 DOORBELL_SOURCE, 0);
2610 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
2614 mqd->cp_hqd_pq_doorbell_control = tmp;
2616 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */
2618 mqd->cp_hqd_pq_rptr = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR);
2620 /* set the vmid for the queue */
2621 mqd->cp_hqd_vmid = 0;
2623 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PERSISTENT_STATE);
2624 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE, PRELOAD_SIZE, 0x53);
2625 mqd->cp_hqd_persistent_state = tmp;
2627 /* set MIN_IB_AVAIL_SIZE */
2628 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_IB_CONTROL);
2629 tmp = REG_SET_FIELD(tmp, CP_HQD_IB_CONTROL, MIN_IB_AVAIL_SIZE, 3);
2630 mqd->cp_hqd_ib_control = tmp;
2632 /* activate the queue */
2633 mqd->cp_hqd_active = 1;
2638 static int gfx_v9_0_kiq_init_register(struct amdgpu_ring *ring)
2640 struct amdgpu_device *adev = ring->adev;
2641 struct v9_mqd *mqd = ring->mqd_ptr;
2644 /* disable wptr polling */
2645 WREG32_FIELD15(GC, 0, CP_PQ_WPTR_POLL_CNTL, EN, 0);
2647 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_BASE_ADDR,
2648 mqd->cp_hqd_eop_base_addr_lo);
2649 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_BASE_ADDR_HI,
2650 mqd->cp_hqd_eop_base_addr_hi);
2652 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
2653 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_CONTROL,
2654 mqd->cp_hqd_eop_control);
2656 /* enable doorbell? */
2657 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL,
2658 mqd->cp_hqd_pq_doorbell_control);
2660 /* disable the queue if it's active */
2661 if (RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1) {
2662 WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, 1);
2663 for (j = 0; j < adev->usec_timeout; j++) {
2664 if (!(RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1))
2668 WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST,
2669 mqd->cp_hqd_dequeue_request);
2670 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR,
2671 mqd->cp_hqd_pq_rptr);
2672 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_LO,
2673 mqd->cp_hqd_pq_wptr_lo);
2674 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_HI,
2675 mqd->cp_hqd_pq_wptr_hi);
2678 /* set the pointer to the MQD */
2679 WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR,
2680 mqd->cp_mqd_base_addr_lo);
2681 WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR_HI,
2682 mqd->cp_mqd_base_addr_hi);
2684 /* set MQD vmid to 0 */
2685 WREG32_SOC15(GC, 0, mmCP_MQD_CONTROL,
2686 mqd->cp_mqd_control);
2688 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
2689 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE,
2690 mqd->cp_hqd_pq_base_lo);
2691 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE_HI,
2692 mqd->cp_hqd_pq_base_hi);
2694 /* set up the HQD, this is similar to CP_RB0_CNTL */
2695 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_CONTROL,
2696 mqd->cp_hqd_pq_control);
2698 /* set the wb address whether it's enabled or not */
2699 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR_REPORT_ADDR,
2700 mqd->cp_hqd_pq_rptr_report_addr_lo);
2701 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
2702 mqd->cp_hqd_pq_rptr_report_addr_hi);
2704 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
2705 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR,
2706 mqd->cp_hqd_pq_wptr_poll_addr_lo);
2707 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI,
2708 mqd->cp_hqd_pq_wptr_poll_addr_hi);
2710 /* enable the doorbell if requested */
2711 if (ring->use_doorbell) {
2712 WREG32_SOC15(GC, 0, mmCP_MEC_DOORBELL_RANGE_LOWER,
2713 (AMDGPU_DOORBELL64_KIQ *2) << 2);
2714 WREG32_SOC15(GC, 0, mmCP_MEC_DOORBELL_RANGE_UPPER,
2715 (AMDGPU_DOORBELL64_USERQUEUE_END * 2) << 2);
2718 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL,
2719 mqd->cp_hqd_pq_doorbell_control);
2721 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */
2722 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_LO,
2723 mqd->cp_hqd_pq_wptr_lo);
2724 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_HI,
2725 mqd->cp_hqd_pq_wptr_hi);
2727 /* set the vmid for the queue */
2728 WREG32_SOC15(GC, 0, mmCP_HQD_VMID, mqd->cp_hqd_vmid);
2730 WREG32_SOC15(GC, 0, mmCP_HQD_PERSISTENT_STATE,
2731 mqd->cp_hqd_persistent_state);
2733 /* activate the queue */
2734 WREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE,
2735 mqd->cp_hqd_active);
2737 if (ring->use_doorbell)
2738 WREG32_FIELD15(GC, 0, CP_PQ_STATUS, DOORBELL_ENABLE, 1);
2743 static int gfx_v9_0_kiq_init_queue(struct amdgpu_ring *ring)
2745 struct amdgpu_device *adev = ring->adev;
2746 struct v9_mqd *mqd = ring->mqd_ptr;
2747 int mqd_idx = AMDGPU_MAX_COMPUTE_RINGS;
2749 gfx_v9_0_kiq_setting(ring);
2751 if (adev->gfx.in_reset) { /* for GPU_RESET case */
2752 /* reset MQD to a clean status */
2753 if (adev->gfx.mec.mqd_backup[mqd_idx])
2754 memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd));
2756 /* reset ring buffer */
2758 amdgpu_ring_clear_ring(ring);
2760 mutex_lock(&adev->srbm_mutex);
2761 soc15_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
2762 gfx_v9_0_kiq_init_register(ring);
2763 soc15_grbm_select(adev, 0, 0, 0, 0);
2764 mutex_unlock(&adev->srbm_mutex);
2766 memset((void *)mqd, 0, sizeof(*mqd));
2767 mutex_lock(&adev->srbm_mutex);
2768 soc15_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
2769 gfx_v9_0_mqd_init(ring);
2770 gfx_v9_0_kiq_init_register(ring);
2771 soc15_grbm_select(adev, 0, 0, 0, 0);
2772 mutex_unlock(&adev->srbm_mutex);
2774 if (adev->gfx.mec.mqd_backup[mqd_idx])
2775 memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
2781 static int gfx_v9_0_kcq_init_queue(struct amdgpu_ring *ring)
2783 struct amdgpu_device *adev = ring->adev;
2784 struct v9_mqd *mqd = ring->mqd_ptr;
2785 int mqd_idx = ring - &adev->gfx.compute_ring[0];
2787 if (!adev->gfx.in_reset && !adev->gfx.in_suspend) {
2788 memset((void *)mqd, 0, sizeof(*mqd));
2789 mutex_lock(&adev->srbm_mutex);
2790 soc15_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
2791 gfx_v9_0_mqd_init(ring);
2792 soc15_grbm_select(adev, 0, 0, 0, 0);
2793 mutex_unlock(&adev->srbm_mutex);
2795 if (adev->gfx.mec.mqd_backup[mqd_idx])
2796 memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
2797 } else if (adev->gfx.in_reset) { /* for GPU_RESET case */
2798 /* reset MQD to a clean status */
2799 if (adev->gfx.mec.mqd_backup[mqd_idx])
2800 memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd));
2802 /* reset ring buffer */
2804 amdgpu_ring_clear_ring(ring);
2806 amdgpu_ring_clear_ring(ring);
2812 static int gfx_v9_0_kiq_resume(struct amdgpu_device *adev)
2814 struct amdgpu_ring *ring = NULL;
2817 gfx_v9_0_cp_compute_enable(adev, true);
2819 ring = &adev->gfx.kiq.ring;
2821 r = amdgpu_bo_reserve(ring->mqd_obj, false);
2822 if (unlikely(r != 0))
2825 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
2827 r = gfx_v9_0_kiq_init_queue(ring);
2828 amdgpu_bo_kunmap(ring->mqd_obj);
2829 ring->mqd_ptr = NULL;
2831 amdgpu_bo_unreserve(ring->mqd_obj);
2835 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
2836 ring = &adev->gfx.compute_ring[i];
2838 r = amdgpu_bo_reserve(ring->mqd_obj, false);
2839 if (unlikely(r != 0))
2841 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
2843 r = gfx_v9_0_kcq_init_queue(ring);
2844 amdgpu_bo_kunmap(ring->mqd_obj);
2845 ring->mqd_ptr = NULL;
2847 amdgpu_bo_unreserve(ring->mqd_obj);
2852 r = gfx_v9_0_kiq_kcq_enable(adev);
2857 static int gfx_v9_0_cp_resume(struct amdgpu_device *adev)
2860 struct amdgpu_ring *ring;
2862 if (!(adev->flags & AMD_IS_APU))
2863 gfx_v9_0_enable_gui_idle_interrupt(adev, false);
2865 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
2866 /* legacy firmware loading */
2867 r = gfx_v9_0_cp_gfx_load_microcode(adev);
2871 r = gfx_v9_0_cp_compute_load_microcode(adev);
2876 r = gfx_v9_0_cp_gfx_resume(adev);
2880 r = gfx_v9_0_kiq_resume(adev);
2884 ring = &adev->gfx.gfx_ring[0];
2885 r = amdgpu_ring_test_ring(ring);
2887 ring->ready = false;
2891 ring = &adev->gfx.kiq.ring;
2893 r = amdgpu_ring_test_ring(ring);
2895 ring->ready = false;
2897 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
2898 ring = &adev->gfx.compute_ring[i];
2901 r = amdgpu_ring_test_ring(ring);
2903 ring->ready = false;
2906 gfx_v9_0_enable_gui_idle_interrupt(adev, true);
2911 static void gfx_v9_0_cp_enable(struct amdgpu_device *adev, bool enable)
2913 gfx_v9_0_cp_gfx_enable(adev, enable);
2914 gfx_v9_0_cp_compute_enable(adev, enable);
2917 static int gfx_v9_0_hw_init(void *handle)
2920 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2922 gfx_v9_0_init_golden_registers(adev);
2924 gfx_v9_0_gpu_init(adev);
2926 r = gfx_v9_0_rlc_resume(adev);
2930 r = gfx_v9_0_cp_resume(adev);
2934 r = gfx_v9_0_ngg_en(adev);
2941 static int gfx_v9_0_hw_fini(void *handle)
2943 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2945 amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0);
2946 amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0);
2947 if (amdgpu_sriov_vf(adev)) {
2948 pr_debug("For SRIOV client, shouldn't do anything.\n");
2951 gfx_v9_0_cp_enable(adev, false);
2952 gfx_v9_0_rlc_stop(adev);
2957 static int gfx_v9_0_suspend(void *handle)
2959 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2961 adev->gfx.in_suspend = true;
2962 return gfx_v9_0_hw_fini(adev);
2965 static int gfx_v9_0_resume(void *handle)
2967 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2970 r = gfx_v9_0_hw_init(adev);
2971 adev->gfx.in_suspend = false;
2975 static bool gfx_v9_0_is_idle(void *handle)
2977 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2979 if (REG_GET_FIELD(RREG32_SOC15(GC, 0, mmGRBM_STATUS),
2980 GRBM_STATUS, GUI_ACTIVE))
2986 static int gfx_v9_0_wait_for_idle(void *handle)
2990 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2992 for (i = 0; i < adev->usec_timeout; i++) {
2993 /* read MC_STATUS */
2994 tmp = RREG32_SOC15(GC, 0, mmGRBM_STATUS) &
2995 GRBM_STATUS__GUI_ACTIVE_MASK;
2997 if (!REG_GET_FIELD(tmp, GRBM_STATUS, GUI_ACTIVE))
3004 static int gfx_v9_0_soft_reset(void *handle)
3006 u32 grbm_soft_reset = 0;
3008 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3011 tmp = RREG32_SOC15(GC, 0, mmGRBM_STATUS);
3012 if (tmp & (GRBM_STATUS__PA_BUSY_MASK | GRBM_STATUS__SC_BUSY_MASK |
3013 GRBM_STATUS__BCI_BUSY_MASK | GRBM_STATUS__SX_BUSY_MASK |
3014 GRBM_STATUS__TA_BUSY_MASK | GRBM_STATUS__VGT_BUSY_MASK |
3015 GRBM_STATUS__DB_BUSY_MASK | GRBM_STATUS__CB_BUSY_MASK |
3016 GRBM_STATUS__GDS_BUSY_MASK | GRBM_STATUS__SPI_BUSY_MASK |
3017 GRBM_STATUS__IA_BUSY_MASK | GRBM_STATUS__IA_BUSY_NO_DMA_MASK)) {
3018 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset,
3019 GRBM_SOFT_RESET, SOFT_RESET_CP, 1);
3020 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset,
3021 GRBM_SOFT_RESET, SOFT_RESET_GFX, 1);
3024 if (tmp & (GRBM_STATUS__CP_BUSY_MASK | GRBM_STATUS__CP_COHERENCY_BUSY_MASK)) {
3025 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset,
3026 GRBM_SOFT_RESET, SOFT_RESET_CP, 1);
3030 tmp = RREG32_SOC15(GC, 0, mmGRBM_STATUS2);
3031 if (REG_GET_FIELD(tmp, GRBM_STATUS2, RLC_BUSY))
3032 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset,
3033 GRBM_SOFT_RESET, SOFT_RESET_RLC, 1);
3036 if (grbm_soft_reset) {
3038 gfx_v9_0_rlc_stop(adev);
3040 /* Disable GFX parsing/prefetching */
3041 gfx_v9_0_cp_gfx_enable(adev, false);
3043 /* Disable MEC parsing/prefetching */
3044 gfx_v9_0_cp_compute_enable(adev, false);
3046 if (grbm_soft_reset) {
3047 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
3048 tmp |= grbm_soft_reset;
3049 dev_info(adev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
3050 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
3051 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
3055 tmp &= ~grbm_soft_reset;
3056 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
3057 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
3060 /* Wait a little for things to settle down */
3066 static uint64_t gfx_v9_0_get_gpu_clock_counter(struct amdgpu_device *adev)
3070 mutex_lock(&adev->gfx.gpu_clock_mutex);
3071 WREG32_SOC15(GC, 0, mmRLC_CAPTURE_GPU_CLOCK_COUNT, 1);
3072 clock = (uint64_t)RREG32_SOC15(GC, 0, mmRLC_GPU_CLOCK_COUNT_LSB) |
3073 ((uint64_t)RREG32_SOC15(GC, 0, mmRLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
3074 mutex_unlock(&adev->gfx.gpu_clock_mutex);
3078 static void gfx_v9_0_ring_emit_gds_switch(struct amdgpu_ring *ring,
3080 uint32_t gds_base, uint32_t gds_size,
3081 uint32_t gws_base, uint32_t gws_size,
3082 uint32_t oa_base, uint32_t oa_size)
3084 gds_base = gds_base >> AMDGPU_GDS_SHIFT;
3085 gds_size = gds_size >> AMDGPU_GDS_SHIFT;
3087 gws_base = gws_base >> AMDGPU_GWS_SHIFT;
3088 gws_size = gws_size >> AMDGPU_GWS_SHIFT;
3090 oa_base = oa_base >> AMDGPU_OA_SHIFT;
3091 oa_size = oa_size >> AMDGPU_OA_SHIFT;
3094 gfx_v9_0_write_data_to_reg(ring, 0, false,
3095 amdgpu_gds_reg_offset[vmid].mem_base,
3099 gfx_v9_0_write_data_to_reg(ring, 0, false,
3100 amdgpu_gds_reg_offset[vmid].mem_size,
3104 gfx_v9_0_write_data_to_reg(ring, 0, false,
3105 amdgpu_gds_reg_offset[vmid].gws,
3106 gws_size << GDS_GWS_VMID0__SIZE__SHIFT | gws_base);
3109 gfx_v9_0_write_data_to_reg(ring, 0, false,
3110 amdgpu_gds_reg_offset[vmid].oa,
3111 (1 << (oa_size + oa_base)) - (1 << oa_base));
3114 static int gfx_v9_0_early_init(void *handle)
3116 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3118 adev->gfx.num_gfx_rings = GFX9_NUM_GFX_RINGS;
3119 adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
3120 gfx_v9_0_set_ring_funcs(adev);
3121 gfx_v9_0_set_irq_funcs(adev);
3122 gfx_v9_0_set_gds_init(adev);
3123 gfx_v9_0_set_rlc_funcs(adev);
3128 static int gfx_v9_0_late_init(void *handle)
3130 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3133 r = amdgpu_irq_get(adev, &adev->gfx.priv_reg_irq, 0);
3137 r = amdgpu_irq_get(adev, &adev->gfx.priv_inst_irq, 0);
3144 static void gfx_v9_0_enter_rlc_safe_mode(struct amdgpu_device *adev)
3146 uint32_t rlc_setting, data;
3149 if (adev->gfx.rlc.in_safe_mode)
3152 /* if RLC is not enabled, do nothing */
3153 rlc_setting = RREG32_SOC15(GC, 0, mmRLC_CNTL);
3154 if (!(rlc_setting & RLC_CNTL__RLC_ENABLE_F32_MASK))
3157 if (adev->cg_flags &
3158 (AMD_CG_SUPPORT_GFX_CGCG | AMD_CG_SUPPORT_GFX_MGCG |
3159 AMD_CG_SUPPORT_GFX_3D_CGCG)) {
3160 data = RLC_SAFE_MODE__CMD_MASK;
3161 data |= (1 << RLC_SAFE_MODE__MESSAGE__SHIFT);
3162 WREG32_SOC15(GC, 0, mmRLC_SAFE_MODE, data);
3164 /* wait for RLC_SAFE_MODE */
3165 for (i = 0; i < adev->usec_timeout; i++) {
3166 if (!REG_GET_FIELD(SOC15_REG_OFFSET(GC, 0, mmRLC_SAFE_MODE), RLC_SAFE_MODE, CMD))
3170 adev->gfx.rlc.in_safe_mode = true;
3174 static void gfx_v9_0_exit_rlc_safe_mode(struct amdgpu_device *adev)
3176 uint32_t rlc_setting, data;
3178 if (!adev->gfx.rlc.in_safe_mode)
3181 /* if RLC is not enabled, do nothing */
3182 rlc_setting = RREG32_SOC15(GC, 0, mmRLC_CNTL);
3183 if (!(rlc_setting & RLC_CNTL__RLC_ENABLE_F32_MASK))
3186 if (adev->cg_flags &
3187 (AMD_CG_SUPPORT_GFX_CGCG | AMD_CG_SUPPORT_GFX_MGCG)) {
3189 * Try to exit safe mode only if it is already in safe
3192 data = RLC_SAFE_MODE__CMD_MASK;
3193 WREG32_SOC15(GC, 0, mmRLC_SAFE_MODE, data);
3194 adev->gfx.rlc.in_safe_mode = false;
3198 static void gfx_v9_0_update_gfx_cg_power_gating(struct amdgpu_device *adev,
3201 /* TODO: double check if we need to perform under safe mdoe */
3202 /* gfx_v9_0_enter_rlc_safe_mode(adev); */
3204 if ((adev->pg_flags & AMD_PG_SUPPORT_GFX_PG) && enable) {
3205 gfx_v9_0_enable_gfx_cg_power_gating(adev, true);
3206 if (adev->pg_flags & AMD_PG_SUPPORT_GFX_PIPELINE)
3207 gfx_v9_0_enable_gfx_pipeline_powergating(adev, true);
3209 gfx_v9_0_enable_gfx_cg_power_gating(adev, false);
3210 gfx_v9_0_enable_gfx_pipeline_powergating(adev, false);
3213 /* gfx_v9_0_exit_rlc_safe_mode(adev); */
3216 static void gfx_v9_0_update_gfx_mg_power_gating(struct amdgpu_device *adev,
3219 /* TODO: double check if we need to perform under safe mode */
3220 /* gfx_v9_0_enter_rlc_safe_mode(adev); */
3222 if ((adev->pg_flags & AMD_PG_SUPPORT_GFX_SMG) && enable)
3223 gfx_v9_0_enable_gfx_static_mg_power_gating(adev, true);
3225 gfx_v9_0_enable_gfx_static_mg_power_gating(adev, false);
3227 if ((adev->pg_flags & AMD_PG_SUPPORT_GFX_DMG) && enable)
3228 gfx_v9_0_enable_gfx_dynamic_mg_power_gating(adev, true);
3230 gfx_v9_0_enable_gfx_dynamic_mg_power_gating(adev, false);
3232 /* gfx_v9_0_exit_rlc_safe_mode(adev); */
3235 static void gfx_v9_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
3240 /* It is disabled by HW by default */
3241 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG)) {
3242 /* 1 - RLC_CGTT_MGCG_OVERRIDE */
3243 def = data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
3244 data &= ~(RLC_CGTT_MGCG_OVERRIDE__CPF_CGTT_SCLK_OVERRIDE_MASK |
3245 RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK |
3246 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK |
3247 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGLS_OVERRIDE_MASK);
3249 /* only for Vega10 & Raven1 */
3250 data |= RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK;
3253 WREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE, data);
3255 /* MGLS is a global flag to control all MGLS in GFX */
3256 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGLS) {
3257 /* 2 - RLC memory Light sleep */
3258 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_RLC_LS) {
3259 def = data = RREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL);
3260 data |= RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK;
3262 WREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL, data);
3264 /* 3 - CP memory Light sleep */
3265 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CP_LS) {
3266 def = data = RREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL);
3267 data |= CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK;
3269 WREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL, data);
3273 /* 1 - MGCG_OVERRIDE */
3274 def = data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
3275 data |= (RLC_CGTT_MGCG_OVERRIDE__CPF_CGTT_SCLK_OVERRIDE_MASK |
3276 RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK |
3277 RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK |
3278 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK |
3279 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGLS_OVERRIDE_MASK);
3281 WREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE, data);
3283 /* 2 - disable MGLS in RLC */
3284 data = RREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL);
3285 if (data & RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK) {
3286 data &= ~RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK;
3287 WREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL, data);
3290 /* 3 - disable MGLS in CP */
3291 data = RREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL);
3292 if (data & CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK) {
3293 data &= ~CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK;
3294 WREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL, data);
3299 static void gfx_v9_0_update_3d_clock_gating(struct amdgpu_device *adev,
3304 adev->gfx.rlc.funcs->enter_safe_mode(adev);
3306 /* Enable 3D CGCG/CGLS */
3307 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG)) {
3308 /* write cmd to clear cgcg/cgls ov */
3309 def = data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
3310 /* unset CGCG override */
3311 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_GFX3D_CG_OVERRIDE_MASK;
3312 /* update CGCG and CGLS override bits */
3314 WREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE, data);
3315 /* enable 3Dcgcg FSM(0x0020003f) */
3316 def = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D);
3317 data = (0x2000 << RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD__SHIFT) |
3318 RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK;
3319 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS)
3320 data |= (0x000F << RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY__SHIFT) |
3321 RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK;
3323 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D, data);
3325 /* set IDLE_POLL_COUNT(0x00900100) */
3326 def = RREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL);
3327 data = (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) |
3328 (0x0090 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT);
3330 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL, data);
3332 /* Disable CGCG/CGLS */
3333 def = data = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D);
3334 /* disable cgcg, cgls should be disabled */
3335 data &= ~(RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK |
3336 RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK);
3337 /* disable cgcg and cgls in FSM */
3339 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D, data);
3342 adev->gfx.rlc.funcs->exit_safe_mode(adev);
3345 static void gfx_v9_0_update_coarse_grain_clock_gating(struct amdgpu_device *adev,
3350 adev->gfx.rlc.funcs->enter_safe_mode(adev);
3352 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG)) {
3353 def = data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
3354 /* unset CGCG override */
3355 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGCG_OVERRIDE_MASK;
3356 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS)
3357 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGLS_OVERRIDE_MASK;
3359 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGLS_OVERRIDE_MASK;
3360 /* update CGCG and CGLS override bits */
3362 WREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE, data);
3364 /* enable cgcg FSM(0x0020003F) */
3365 def = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL);
3366 data = (0x2000 << RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD__SHIFT) |
3367 RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK;
3368 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS)
3369 data |= (0x000F << RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY__SHIFT) |
3370 RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK;
3372 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL, data);
3374 /* set IDLE_POLL_COUNT(0x00900100) */
3375 def = RREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL);
3376 data = (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) |
3377 (0x0090 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT);
3379 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL, data);
3381 def = data = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL);
3382 /* reset CGCG/CGLS bits */
3383 data &= ~(RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK | RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK);
3384 /* disable cgcg and cgls in FSM */
3386 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL, data);
3389 adev->gfx.rlc.funcs->exit_safe_mode(adev);
3392 static int gfx_v9_0_update_gfx_clock_gating(struct amdgpu_device *adev,
3396 /* CGCG/CGLS should be enabled after MGCG/MGLS
3397 * === MGCG + MGLS ===
3399 gfx_v9_0_update_medium_grain_clock_gating(adev, enable);
3400 /* === CGCG /CGLS for GFX 3D Only === */
3401 gfx_v9_0_update_3d_clock_gating(adev, enable);
3402 /* === CGCG + CGLS === */
3403 gfx_v9_0_update_coarse_grain_clock_gating(adev, enable);
3405 /* CGCG/CGLS should be disabled before MGCG/MGLS
3406 * === CGCG + CGLS ===
3408 gfx_v9_0_update_coarse_grain_clock_gating(adev, enable);
3409 /* === CGCG /CGLS for GFX 3D Only === */
3410 gfx_v9_0_update_3d_clock_gating(adev, enable);
3411 /* === MGCG + MGLS === */
3412 gfx_v9_0_update_medium_grain_clock_gating(adev, enable);
3417 static const struct amdgpu_rlc_funcs gfx_v9_0_rlc_funcs = {
3418 .enter_safe_mode = gfx_v9_0_enter_rlc_safe_mode,
3419 .exit_safe_mode = gfx_v9_0_exit_rlc_safe_mode
3422 static int gfx_v9_0_set_powergating_state(void *handle,
3423 enum amd_powergating_state state)
3425 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3426 bool enable = (state == AMD_PG_STATE_GATE) ? true : false;
3428 switch (adev->asic_type) {
3430 if (adev->pg_flags & AMD_PG_SUPPORT_RLC_SMU_HS) {
3431 gfx_v9_0_enable_sck_slow_down_on_power_up(adev, true);
3432 gfx_v9_0_enable_sck_slow_down_on_power_down(adev, true);
3434 gfx_v9_0_enable_sck_slow_down_on_power_up(adev, false);
3435 gfx_v9_0_enable_sck_slow_down_on_power_down(adev, false);
3438 if (adev->pg_flags & AMD_PG_SUPPORT_CP)
3439 gfx_v9_0_enable_cp_power_gating(adev, true);
3441 gfx_v9_0_enable_cp_power_gating(adev, false);
3443 /* update gfx cgpg state */
3444 gfx_v9_0_update_gfx_cg_power_gating(adev, enable);
3446 /* update mgcg state */
3447 gfx_v9_0_update_gfx_mg_power_gating(adev, enable);
3456 static int gfx_v9_0_set_clockgating_state(void *handle,
3457 enum amd_clockgating_state state)
3459 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3461 if (amdgpu_sriov_vf(adev))
3464 switch (adev->asic_type) {
3467 gfx_v9_0_update_gfx_clock_gating(adev,
3468 state == AMD_CG_STATE_GATE ? true : false);
3476 static void gfx_v9_0_get_clockgating_state(void *handle, u32 *flags)
3478 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3481 if (amdgpu_sriov_vf(adev))
3484 /* AMD_CG_SUPPORT_GFX_MGCG */
3485 data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
3486 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK))
3487 *flags |= AMD_CG_SUPPORT_GFX_MGCG;
3489 /* AMD_CG_SUPPORT_GFX_CGCG */
3490 data = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL);
3491 if (data & RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK)
3492 *flags |= AMD_CG_SUPPORT_GFX_CGCG;
3494 /* AMD_CG_SUPPORT_GFX_CGLS */
3495 if (data & RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK)
3496 *flags |= AMD_CG_SUPPORT_GFX_CGLS;
3498 /* AMD_CG_SUPPORT_GFX_RLC_LS */
3499 data = RREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL);
3500 if (data & RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK)
3501 *flags |= AMD_CG_SUPPORT_GFX_RLC_LS | AMD_CG_SUPPORT_GFX_MGLS;
3503 /* AMD_CG_SUPPORT_GFX_CP_LS */
3504 data = RREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL);
3505 if (data & CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK)
3506 *flags |= AMD_CG_SUPPORT_GFX_CP_LS | AMD_CG_SUPPORT_GFX_MGLS;
3508 /* AMD_CG_SUPPORT_GFX_3D_CGCG */
3509 data = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D);
3510 if (data & RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK)
3511 *flags |= AMD_CG_SUPPORT_GFX_3D_CGCG;
3513 /* AMD_CG_SUPPORT_GFX_3D_CGLS */
3514 if (data & RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK)
3515 *flags |= AMD_CG_SUPPORT_GFX_3D_CGLS;
3518 static u64 gfx_v9_0_ring_get_rptr_gfx(struct amdgpu_ring *ring)
3520 return ring->adev->wb.wb[ring->rptr_offs]; /* gfx9 is 32bit rptr*/
3523 static u64 gfx_v9_0_ring_get_wptr_gfx(struct amdgpu_ring *ring)
3525 struct amdgpu_device *adev = ring->adev;
3528 /* XXX check if swapping is necessary on BE */
3529 if (ring->use_doorbell) {
3530 wptr = atomic64_read((atomic64_t *)&adev->wb.wb[ring->wptr_offs]);
3532 wptr = RREG32_SOC15(GC, 0, mmCP_RB0_WPTR);
3533 wptr += (u64)RREG32_SOC15(GC, 0, mmCP_RB0_WPTR_HI) << 32;
3539 static void gfx_v9_0_ring_set_wptr_gfx(struct amdgpu_ring *ring)
3541 struct amdgpu_device *adev = ring->adev;
3543 if (ring->use_doorbell) {
3544 /* XXX check if swapping is necessary on BE */
3545 atomic64_set((atomic64_t*)&adev->wb.wb[ring->wptr_offs], ring->wptr);
3546 WDOORBELL64(ring->doorbell_index, ring->wptr);
3548 WREG32_SOC15(GC, 0, mmCP_RB0_WPTR, lower_32_bits(ring->wptr));
3549 WREG32_SOC15(GC, 0, mmCP_RB0_WPTR_HI, upper_32_bits(ring->wptr));
3553 static void gfx_v9_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
3555 u32 ref_and_mask, reg_mem_engine;
3556 struct nbio_hdp_flush_reg *nbio_hf_reg;
3558 if (ring->adev->asic_type == CHIP_VEGA10)
3559 nbio_hf_reg = &nbio_v6_1_hdp_flush_reg;
3561 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE) {
3564 ref_and_mask = nbio_hf_reg->ref_and_mask_cp2 << ring->pipe;
3567 ref_and_mask = nbio_hf_reg->ref_and_mask_cp6 << ring->pipe;
3574 ref_and_mask = nbio_hf_reg->ref_and_mask_cp0;
3575 reg_mem_engine = 1; /* pfp */
3578 gfx_v9_0_wait_reg_mem(ring, reg_mem_engine, 0, 1,
3579 nbio_hf_reg->hdp_flush_req_offset,
3580 nbio_hf_reg->hdp_flush_done_offset,
3581 ref_and_mask, ref_and_mask, 0x20);
3584 static void gfx_v9_0_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
3586 gfx_v9_0_write_data_to_reg(ring, 0, true,
3587 SOC15_REG_OFFSET(HDP, 0, mmHDP_DEBUG0), 1);
3590 static void gfx_v9_0_ring_emit_ib_gfx(struct amdgpu_ring *ring,
3591 struct amdgpu_ib *ib,
3592 unsigned vm_id, bool ctx_switch)
3594 u32 header, control = 0;
3596 if (ib->flags & AMDGPU_IB_FLAG_CE)
3597 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
3599 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
3601 control |= ib->length_dw | (vm_id << 24);
3603 if (amdgpu_sriov_vf(ring->adev) && (ib->flags & AMDGPU_IB_FLAG_PREEMPT)) {
3604 control |= INDIRECT_BUFFER_PRE_ENB(1);
3606 if (!(ib->flags & AMDGPU_IB_FLAG_CE))
3607 gfx_v9_0_ring_emit_de_meta(ring);
3610 amdgpu_ring_write(ring, header);
3611 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */
3612 amdgpu_ring_write(ring,
3616 lower_32_bits(ib->gpu_addr));
3617 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
3618 amdgpu_ring_write(ring, control);
3621 static void gfx_v9_0_ring_emit_ib_compute(struct amdgpu_ring *ring,
3622 struct amdgpu_ib *ib,
3623 unsigned vm_id, bool ctx_switch)
3625 u32 control = INDIRECT_BUFFER_VALID | ib->length_dw | (vm_id << 24);
3627 amdgpu_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2));
3628 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */
3629 amdgpu_ring_write(ring,
3633 lower_32_bits(ib->gpu_addr));
3634 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
3635 amdgpu_ring_write(ring, control);
3638 static void gfx_v9_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr,
3639 u64 seq, unsigned flags)
3641 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
3642 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
3644 /* RELEASE_MEM - flush caches, send int */
3645 amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6));
3646 amdgpu_ring_write(ring, (EOP_TCL1_ACTION_EN |
3648 EOP_TC_WB_ACTION_EN |
3649 EOP_TC_MD_ACTION_EN |
3650 EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
3652 amdgpu_ring_write(ring, DATA_SEL(write64bit ? 2 : 1) | INT_SEL(int_sel ? 2 : 0));
3655 * the address should be Qword aligned if 64bit write, Dword
3656 * aligned if only send 32bit data low (discard data high)
3662 amdgpu_ring_write(ring, lower_32_bits(addr));
3663 amdgpu_ring_write(ring, upper_32_bits(addr));
3664 amdgpu_ring_write(ring, lower_32_bits(seq));
3665 amdgpu_ring_write(ring, upper_32_bits(seq));
3666 amdgpu_ring_write(ring, 0);
3669 static void gfx_v9_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
3671 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX);
3672 uint32_t seq = ring->fence_drv.sync_seq;
3673 uint64_t addr = ring->fence_drv.gpu_addr;
3675 gfx_v9_0_wait_reg_mem(ring, usepfp, 1, 0,
3676 lower_32_bits(addr), upper_32_bits(addr),
3677 seq, 0xffffffff, 4);
3680 static void gfx_v9_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
3681 unsigned vm_id, uint64_t pd_addr)
3683 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub];
3684 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX);
3685 uint32_t req = ring->adev->gart.gart_funcs->get_invalidate_req(vm_id);
3686 unsigned eng = ring->vm_inv_eng;
3688 pd_addr = amdgpu_gart_get_vm_pde(ring->adev, pd_addr);
3689 pd_addr |= AMDGPU_PTE_VALID;
3691 gfx_v9_0_write_data_to_reg(ring, usepfp, true,
3692 hub->ctx0_ptb_addr_lo32 + (2 * vm_id),
3693 lower_32_bits(pd_addr));
3695 gfx_v9_0_write_data_to_reg(ring, usepfp, true,
3696 hub->ctx0_ptb_addr_hi32 + (2 * vm_id),
3697 upper_32_bits(pd_addr));
3699 gfx_v9_0_write_data_to_reg(ring, usepfp, true,
3700 hub->vm_inv_eng0_req + eng, req);
3702 /* wait for the invalidate to complete */
3703 gfx_v9_0_wait_reg_mem(ring, 0, 0, 0, hub->vm_inv_eng0_ack +
3704 eng, 0, 1 << vm_id, 1 << vm_id, 0x20);
3706 /* compute doesn't have PFP */
3708 /* sync PFP to ME, otherwise we might get invalid PFP reads */
3709 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
3710 amdgpu_ring_write(ring, 0x0);
3714 static u64 gfx_v9_0_ring_get_rptr_compute(struct amdgpu_ring *ring)
3716 return ring->adev->wb.wb[ring->rptr_offs]; /* gfx9 hardware is 32bit rptr */
3719 static u64 gfx_v9_0_ring_get_wptr_compute(struct amdgpu_ring *ring)
3723 /* XXX check if swapping is necessary on BE */
3724 if (ring->use_doorbell)
3725 wptr = atomic64_read((atomic64_t *)&ring->adev->wb.wb[ring->wptr_offs]);
3731 static void gfx_v9_0_ring_set_wptr_compute(struct amdgpu_ring *ring)
3733 struct amdgpu_device *adev = ring->adev;
3735 /* XXX check if swapping is necessary on BE */
3736 if (ring->use_doorbell) {
3737 atomic64_set((atomic64_t*)&adev->wb.wb[ring->wptr_offs], ring->wptr);
3738 WDOORBELL64(ring->doorbell_index, ring->wptr);
3740 BUG(); /* only DOORBELL method supported on gfx9 now */
3744 static void gfx_v9_0_ring_emit_fence_kiq(struct amdgpu_ring *ring, u64 addr,
3745 u64 seq, unsigned int flags)
3747 /* we only allocate 32bit for each seq wb address */
3748 BUG_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
3750 /* write fence seq to the "addr" */
3751 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3752 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3753 WRITE_DATA_DST_SEL(5) | WR_CONFIRM));
3754 amdgpu_ring_write(ring, lower_32_bits(addr));
3755 amdgpu_ring_write(ring, upper_32_bits(addr));
3756 amdgpu_ring_write(ring, lower_32_bits(seq));
3758 if (flags & AMDGPU_FENCE_FLAG_INT) {
3759 /* set register to trigger INT */
3760 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3761 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3762 WRITE_DATA_DST_SEL(0) | WR_CONFIRM));
3763 amdgpu_ring_write(ring, SOC15_REG_OFFSET(GC, 0, mmCPC_INT_STATUS));
3764 amdgpu_ring_write(ring, 0);
3765 amdgpu_ring_write(ring, 0x20000000); /* src_id is 178 */
3769 static void gfx_v9_ring_emit_sb(struct amdgpu_ring *ring)
3771 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3772 amdgpu_ring_write(ring, 0);
3775 static void gfx_v9_0_ring_emit_ce_meta(struct amdgpu_ring *ring)
3777 static struct v9_ce_ib_state ce_payload = {0};
3781 cnt = (sizeof(ce_payload) >> 2) + 4 - 2;
3782 csa_addr = AMDGPU_VA_RESERVED_SIZE - 2 * 4096;
3784 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, cnt));
3785 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(2) |
3786 WRITE_DATA_DST_SEL(8) |
3788 WRITE_DATA_CACHE_POLICY(0));
3789 amdgpu_ring_write(ring, lower_32_bits(csa_addr + offsetof(struct v9_gfx_meta_data, ce_payload)));
3790 amdgpu_ring_write(ring, upper_32_bits(csa_addr + offsetof(struct v9_gfx_meta_data, ce_payload)));
3791 amdgpu_ring_write_multiple(ring, (void *)&ce_payload, sizeof(ce_payload) >> 2);
3794 static void gfx_v9_0_ring_emit_de_meta(struct amdgpu_ring *ring)
3796 static struct v9_de_ib_state de_payload = {0};
3797 uint64_t csa_addr, gds_addr;
3800 csa_addr = AMDGPU_VA_RESERVED_SIZE - 2 * 4096;
3801 gds_addr = csa_addr + 4096;
3802 de_payload.gds_backup_addrlo = lower_32_bits(gds_addr);
3803 de_payload.gds_backup_addrhi = upper_32_bits(gds_addr);
3805 cnt = (sizeof(de_payload) >> 2) + 4 - 2;
3806 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, cnt));
3807 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
3808 WRITE_DATA_DST_SEL(8) |
3810 WRITE_DATA_CACHE_POLICY(0));
3811 amdgpu_ring_write(ring, lower_32_bits(csa_addr + offsetof(struct v9_gfx_meta_data, de_payload)));
3812 amdgpu_ring_write(ring, upper_32_bits(csa_addr + offsetof(struct v9_gfx_meta_data, de_payload)));
3813 amdgpu_ring_write_multiple(ring, (void *)&de_payload, sizeof(de_payload) >> 2);
3816 static void gfx_v9_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags)
3820 if (amdgpu_sriov_vf(ring->adev))
3821 gfx_v9_0_ring_emit_ce_meta(ring);
3823 dw2 |= 0x80000000; /* set load_enable otherwise this package is just NOPs */
3824 if (flags & AMDGPU_HAVE_CTX_SWITCH) {
3825 /* set load_global_config & load_global_uconfig */
3827 /* set load_cs_sh_regs */
3829 /* set load_per_context_state & load_gfx_sh_regs for GFX */
3832 /* set load_ce_ram if preamble presented */
3833 if (AMDGPU_PREAMBLE_IB_PRESENT & flags)
3836 /* still load_ce_ram if this is the first time preamble presented
3837 * although there is no context switch happens.
3839 if (AMDGPU_PREAMBLE_IB_PRESENT_FIRST & flags)
3843 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
3844 amdgpu_ring_write(ring, dw2);
3845 amdgpu_ring_write(ring, 0);
3848 static unsigned gfx_v9_0_ring_emit_init_cond_exec(struct amdgpu_ring *ring)
3851 amdgpu_ring_write(ring, PACKET3(PACKET3_COND_EXEC, 3));
3852 amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
3853 amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
3854 amdgpu_ring_write(ring, 0); /* discard following DWs if *cond_exec_gpu_addr==0 */
3855 ret = ring->wptr & ring->buf_mask;
3856 amdgpu_ring_write(ring, 0x55aa55aa); /* patch dummy value later */
3860 static void gfx_v9_0_ring_emit_patch_cond_exec(struct amdgpu_ring *ring, unsigned offset)
3863 BUG_ON(offset > ring->buf_mask);
3864 BUG_ON(ring->ring[offset] != 0x55aa55aa);
3866 cur = (ring->wptr & ring->buf_mask) - 1;
3867 if (likely(cur > offset))
3868 ring->ring[offset] = cur - offset;
3870 ring->ring[offset] = (ring->ring_size>>2) - offset + cur;
3873 static void gfx_v9_0_ring_emit_tmz(struct amdgpu_ring *ring, bool start)
3875 amdgpu_ring_write(ring, PACKET3(PACKET3_FRAME_CONTROL, 0));
3876 amdgpu_ring_write(ring, FRAME_CMD(start ? 0 : 1)); /* frame_end */
3879 static void gfx_v9_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg)
3881 struct amdgpu_device *adev = ring->adev;
3883 amdgpu_ring_write(ring, PACKET3(PACKET3_COPY_DATA, 4));
3884 amdgpu_ring_write(ring, 0 | /* src: register*/
3885 (5 << 8) | /* dst: memory */
3886 (1 << 20)); /* write confirm */
3887 amdgpu_ring_write(ring, reg);
3888 amdgpu_ring_write(ring, 0);
3889 amdgpu_ring_write(ring, lower_32_bits(adev->wb.gpu_addr +
3890 adev->virt.reg_val_offs * 4));
3891 amdgpu_ring_write(ring, upper_32_bits(adev->wb.gpu_addr +
3892 adev->virt.reg_val_offs * 4));
3895 static void gfx_v9_0_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg,
3898 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3899 amdgpu_ring_write(ring, (1 << 16)); /* no inc addr */
3900 amdgpu_ring_write(ring, reg);
3901 amdgpu_ring_write(ring, 0);
3902 amdgpu_ring_write(ring, val);
3905 static void gfx_v9_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev,
3906 enum amdgpu_interrupt_state state)
3909 case AMDGPU_IRQ_STATE_DISABLE:
3910 case AMDGPU_IRQ_STATE_ENABLE:
3911 WREG32_FIELD15(GC, 0, CP_INT_CNTL_RING0,
3912 TIME_STAMP_INT_ENABLE,
3913 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
3920 static void gfx_v9_0_set_compute_eop_interrupt_state(struct amdgpu_device *adev,
3922 enum amdgpu_interrupt_state state)
3924 u32 mec_int_cntl, mec_int_cntl_reg;
3927 * amdgpu controls only the first MEC. That's why this function only
3928 * handles the setting of interrupts for this specific MEC. All other
3929 * pipes' interrupts are set by amdkfd.
3935 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, mmCP_ME1_PIPE0_INT_CNTL);
3938 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, mmCP_ME1_PIPE1_INT_CNTL);
3941 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, mmCP_ME1_PIPE2_INT_CNTL);
3944 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, mmCP_ME1_PIPE3_INT_CNTL);
3947 DRM_DEBUG("invalid pipe %d\n", pipe);
3951 DRM_DEBUG("invalid me %d\n", me);
3956 case AMDGPU_IRQ_STATE_DISABLE:
3957 mec_int_cntl = RREG32(mec_int_cntl_reg);
3958 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL,
3959 TIME_STAMP_INT_ENABLE, 0);
3960 WREG32(mec_int_cntl_reg, mec_int_cntl);
3962 case AMDGPU_IRQ_STATE_ENABLE:
3963 mec_int_cntl = RREG32(mec_int_cntl_reg);
3964 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL,
3965 TIME_STAMP_INT_ENABLE, 1);
3966 WREG32(mec_int_cntl_reg, mec_int_cntl);
3973 static int gfx_v9_0_set_priv_reg_fault_state(struct amdgpu_device *adev,
3974 struct amdgpu_irq_src *source,
3976 enum amdgpu_interrupt_state state)
3979 case AMDGPU_IRQ_STATE_DISABLE:
3980 case AMDGPU_IRQ_STATE_ENABLE:
3981 WREG32_FIELD15(GC, 0, CP_INT_CNTL_RING0,
3982 PRIV_REG_INT_ENABLE,
3983 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
3992 static int gfx_v9_0_set_priv_inst_fault_state(struct amdgpu_device *adev,
3993 struct amdgpu_irq_src *source,
3995 enum amdgpu_interrupt_state state)
3998 case AMDGPU_IRQ_STATE_DISABLE:
3999 case AMDGPU_IRQ_STATE_ENABLE:
4000 WREG32_FIELD15(GC, 0, CP_INT_CNTL_RING0,
4001 PRIV_INSTR_INT_ENABLE,
4002 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
4010 static int gfx_v9_0_set_eop_interrupt_state(struct amdgpu_device *adev,
4011 struct amdgpu_irq_src *src,
4013 enum amdgpu_interrupt_state state)
4016 case AMDGPU_CP_IRQ_GFX_EOP:
4017 gfx_v9_0_set_gfx_eop_interrupt_state(adev, state);
4019 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP:
4020 gfx_v9_0_set_compute_eop_interrupt_state(adev, 1, 0, state);
4022 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE1_EOP:
4023 gfx_v9_0_set_compute_eop_interrupt_state(adev, 1, 1, state);
4025 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE2_EOP:
4026 gfx_v9_0_set_compute_eop_interrupt_state(adev, 1, 2, state);
4028 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE3_EOP:
4029 gfx_v9_0_set_compute_eop_interrupt_state(adev, 1, 3, state);
4031 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE0_EOP:
4032 gfx_v9_0_set_compute_eop_interrupt_state(adev, 2, 0, state);
4034 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE1_EOP:
4035 gfx_v9_0_set_compute_eop_interrupt_state(adev, 2, 1, state);
4037 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE2_EOP:
4038 gfx_v9_0_set_compute_eop_interrupt_state(adev, 2, 2, state);
4040 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE3_EOP:
4041 gfx_v9_0_set_compute_eop_interrupt_state(adev, 2, 3, state);
4049 static int gfx_v9_0_eop_irq(struct amdgpu_device *adev,
4050 struct amdgpu_irq_src *source,
4051 struct amdgpu_iv_entry *entry)
4054 u8 me_id, pipe_id, queue_id;
4055 struct amdgpu_ring *ring;
4057 DRM_DEBUG("IH: CP EOP\n");
4058 me_id = (entry->ring_id & 0x0c) >> 2;
4059 pipe_id = (entry->ring_id & 0x03) >> 0;
4060 queue_id = (entry->ring_id & 0x70) >> 4;
4064 amdgpu_fence_process(&adev->gfx.gfx_ring[0]);
4068 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
4069 ring = &adev->gfx.compute_ring[i];
4070 /* Per-queue interrupt is supported for MEC starting from VI.
4071 * The interrupt can only be enabled/disabled per pipe instead of per queue.
4073 if ((ring->me == me_id) && (ring->pipe == pipe_id) && (ring->queue == queue_id))
4074 amdgpu_fence_process(ring);
4081 static int gfx_v9_0_priv_reg_irq(struct amdgpu_device *adev,
4082 struct amdgpu_irq_src *source,
4083 struct amdgpu_iv_entry *entry)
4085 DRM_ERROR("Illegal register access in command stream\n");
4086 schedule_work(&adev->reset_work);
4090 static int gfx_v9_0_priv_inst_irq(struct amdgpu_device *adev,
4091 struct amdgpu_irq_src *source,
4092 struct amdgpu_iv_entry *entry)
4094 DRM_ERROR("Illegal instruction in command stream\n");
4095 schedule_work(&adev->reset_work);
4099 static int gfx_v9_0_kiq_set_interrupt_state(struct amdgpu_device *adev,
4100 struct amdgpu_irq_src *src,
4102 enum amdgpu_interrupt_state state)
4104 uint32_t tmp, target;
4105 struct amdgpu_ring *ring = &(adev->gfx.kiq.ring);
4108 target = SOC15_REG_OFFSET(GC, 0, mmCP_ME1_PIPE0_INT_CNTL);
4110 target = SOC15_REG_OFFSET(GC, 0, mmCP_ME2_PIPE0_INT_CNTL);
4111 target += ring->pipe;
4114 case AMDGPU_CP_KIQ_IRQ_DRIVER0:
4115 if (state == AMDGPU_IRQ_STATE_DISABLE) {
4116 tmp = RREG32_SOC15(GC, 0, mmCPC_INT_CNTL);
4117 tmp = REG_SET_FIELD(tmp, CPC_INT_CNTL,
4118 GENERIC2_INT_ENABLE, 0);
4119 WREG32_SOC15(GC, 0, mmCPC_INT_CNTL, tmp);
4121 tmp = RREG32(target);
4122 tmp = REG_SET_FIELD(tmp, CP_ME2_PIPE0_INT_CNTL,
4123 GENERIC2_INT_ENABLE, 0);
4124 WREG32(target, tmp);
4126 tmp = RREG32_SOC15(GC, 0, mmCPC_INT_CNTL);
4127 tmp = REG_SET_FIELD(tmp, CPC_INT_CNTL,
4128 GENERIC2_INT_ENABLE, 1);
4129 WREG32_SOC15(GC, 0, mmCPC_INT_CNTL, tmp);
4131 tmp = RREG32(target);
4132 tmp = REG_SET_FIELD(tmp, CP_ME2_PIPE0_INT_CNTL,
4133 GENERIC2_INT_ENABLE, 1);
4134 WREG32(target, tmp);
4138 BUG(); /* kiq only support GENERIC2_INT now */
4144 static int gfx_v9_0_kiq_irq(struct amdgpu_device *adev,
4145 struct amdgpu_irq_src *source,
4146 struct amdgpu_iv_entry *entry)
4148 u8 me_id, pipe_id, queue_id;
4149 struct amdgpu_ring *ring = &(adev->gfx.kiq.ring);
4151 me_id = (entry->ring_id & 0x0c) >> 2;
4152 pipe_id = (entry->ring_id & 0x03) >> 0;
4153 queue_id = (entry->ring_id & 0x70) >> 4;
4154 DRM_DEBUG("IH: CPC GENERIC2_INT, me:%d, pipe:%d, queue:%d\n",
4155 me_id, pipe_id, queue_id);
4157 amdgpu_fence_process(ring);
4161 const struct amd_ip_funcs gfx_v9_0_ip_funcs = {
4163 .early_init = gfx_v9_0_early_init,
4164 .late_init = gfx_v9_0_late_init,
4165 .sw_init = gfx_v9_0_sw_init,
4166 .sw_fini = gfx_v9_0_sw_fini,
4167 .hw_init = gfx_v9_0_hw_init,
4168 .hw_fini = gfx_v9_0_hw_fini,
4169 .suspend = gfx_v9_0_suspend,
4170 .resume = gfx_v9_0_resume,
4171 .is_idle = gfx_v9_0_is_idle,
4172 .wait_for_idle = gfx_v9_0_wait_for_idle,
4173 .soft_reset = gfx_v9_0_soft_reset,
4174 .set_clockgating_state = gfx_v9_0_set_clockgating_state,
4175 .set_powergating_state = gfx_v9_0_set_powergating_state,
4176 .get_clockgating_state = gfx_v9_0_get_clockgating_state,
4179 static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_gfx = {
4180 .type = AMDGPU_RING_TYPE_GFX,
4182 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
4183 .support_64bit_ptrs = true,
4184 .vmhub = AMDGPU_GFXHUB,
4185 .get_rptr = gfx_v9_0_ring_get_rptr_gfx,
4186 .get_wptr = gfx_v9_0_ring_get_wptr_gfx,
4187 .set_wptr = gfx_v9_0_ring_set_wptr_gfx,
4188 .emit_frame_size = /* totally 242 maximum if 16 IBs */
4190 7 + /* PIPELINE_SYNC */
4192 8 + /* FENCE for VM_FLUSH */
4193 20 + /* GDS switch */
4194 4 + /* double SWITCH_BUFFER,
4195 the first COND_EXEC jump to the place just
4196 prior to this double SWITCH_BUFFER */
4204 8 + 8 + /* FENCE x2 */
4205 2, /* SWITCH_BUFFER */
4206 .emit_ib_size = 4, /* gfx_v9_0_ring_emit_ib_gfx */
4207 .emit_ib = gfx_v9_0_ring_emit_ib_gfx,
4208 .emit_fence = gfx_v9_0_ring_emit_fence,
4209 .emit_pipeline_sync = gfx_v9_0_ring_emit_pipeline_sync,
4210 .emit_vm_flush = gfx_v9_0_ring_emit_vm_flush,
4211 .emit_gds_switch = gfx_v9_0_ring_emit_gds_switch,
4212 .emit_hdp_flush = gfx_v9_0_ring_emit_hdp_flush,
4213 .emit_hdp_invalidate = gfx_v9_0_ring_emit_hdp_invalidate,
4214 .test_ring = gfx_v9_0_ring_test_ring,
4215 .test_ib = gfx_v9_0_ring_test_ib,
4216 .insert_nop = amdgpu_ring_insert_nop,
4217 .pad_ib = amdgpu_ring_generic_pad_ib,
4218 .emit_switch_buffer = gfx_v9_ring_emit_sb,
4219 .emit_cntxcntl = gfx_v9_ring_emit_cntxcntl,
4220 .init_cond_exec = gfx_v9_0_ring_emit_init_cond_exec,
4221 .patch_cond_exec = gfx_v9_0_ring_emit_patch_cond_exec,
4222 .emit_tmz = gfx_v9_0_ring_emit_tmz,
4225 static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_compute = {
4226 .type = AMDGPU_RING_TYPE_COMPUTE,
4228 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
4229 .support_64bit_ptrs = true,
4230 .vmhub = AMDGPU_GFXHUB,
4231 .get_rptr = gfx_v9_0_ring_get_rptr_compute,
4232 .get_wptr = gfx_v9_0_ring_get_wptr_compute,
4233 .set_wptr = gfx_v9_0_ring_set_wptr_compute,
4235 20 + /* gfx_v9_0_ring_emit_gds_switch */
4236 7 + /* gfx_v9_0_ring_emit_hdp_flush */
4237 5 + /* gfx_v9_0_ring_emit_hdp_invalidate */
4238 7 + /* gfx_v9_0_ring_emit_pipeline_sync */
4239 24 + /* gfx_v9_0_ring_emit_vm_flush */
4240 8 + 8 + 8, /* gfx_v9_0_ring_emit_fence x3 for user fence, vm fence */
4241 .emit_ib_size = 4, /* gfx_v9_0_ring_emit_ib_compute */
4242 .emit_ib = gfx_v9_0_ring_emit_ib_compute,
4243 .emit_fence = gfx_v9_0_ring_emit_fence,
4244 .emit_pipeline_sync = gfx_v9_0_ring_emit_pipeline_sync,
4245 .emit_vm_flush = gfx_v9_0_ring_emit_vm_flush,
4246 .emit_gds_switch = gfx_v9_0_ring_emit_gds_switch,
4247 .emit_hdp_flush = gfx_v9_0_ring_emit_hdp_flush,
4248 .emit_hdp_invalidate = gfx_v9_0_ring_emit_hdp_invalidate,
4249 .test_ring = gfx_v9_0_ring_test_ring,
4250 .test_ib = gfx_v9_0_ring_test_ib,
4251 .insert_nop = amdgpu_ring_insert_nop,
4252 .pad_ib = amdgpu_ring_generic_pad_ib,
4255 static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_kiq = {
4256 .type = AMDGPU_RING_TYPE_KIQ,
4258 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
4259 .support_64bit_ptrs = true,
4260 .vmhub = AMDGPU_GFXHUB,
4261 .get_rptr = gfx_v9_0_ring_get_rptr_compute,
4262 .get_wptr = gfx_v9_0_ring_get_wptr_compute,
4263 .set_wptr = gfx_v9_0_ring_set_wptr_compute,
4265 20 + /* gfx_v9_0_ring_emit_gds_switch */
4266 7 + /* gfx_v9_0_ring_emit_hdp_flush */
4267 5 + /* gfx_v9_0_ring_emit_hdp_invalidate */
4268 7 + /* gfx_v9_0_ring_emit_pipeline_sync */
4269 24 + /* gfx_v9_0_ring_emit_vm_flush */
4270 8 + 8 + 8, /* gfx_v9_0_ring_emit_fence_kiq x3 for user fence, vm fence */
4271 .emit_ib_size = 4, /* gfx_v9_0_ring_emit_ib_compute */
4272 .emit_ib = gfx_v9_0_ring_emit_ib_compute,
4273 .emit_fence = gfx_v9_0_ring_emit_fence_kiq,
4274 .test_ring = gfx_v9_0_ring_test_ring,
4275 .test_ib = gfx_v9_0_ring_test_ib,
4276 .insert_nop = amdgpu_ring_insert_nop,
4277 .pad_ib = amdgpu_ring_generic_pad_ib,
4278 .emit_rreg = gfx_v9_0_ring_emit_rreg,
4279 .emit_wreg = gfx_v9_0_ring_emit_wreg,
4282 static void gfx_v9_0_set_ring_funcs(struct amdgpu_device *adev)
4286 adev->gfx.kiq.ring.funcs = &gfx_v9_0_ring_funcs_kiq;
4288 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
4289 adev->gfx.gfx_ring[i].funcs = &gfx_v9_0_ring_funcs_gfx;
4291 for (i = 0; i < adev->gfx.num_compute_rings; i++)
4292 adev->gfx.compute_ring[i].funcs = &gfx_v9_0_ring_funcs_compute;
4295 static const struct amdgpu_irq_src_funcs gfx_v9_0_kiq_irq_funcs = {
4296 .set = gfx_v9_0_kiq_set_interrupt_state,
4297 .process = gfx_v9_0_kiq_irq,
4300 static const struct amdgpu_irq_src_funcs gfx_v9_0_eop_irq_funcs = {
4301 .set = gfx_v9_0_set_eop_interrupt_state,
4302 .process = gfx_v9_0_eop_irq,
4305 static const struct amdgpu_irq_src_funcs gfx_v9_0_priv_reg_irq_funcs = {
4306 .set = gfx_v9_0_set_priv_reg_fault_state,
4307 .process = gfx_v9_0_priv_reg_irq,
4310 static const struct amdgpu_irq_src_funcs gfx_v9_0_priv_inst_irq_funcs = {
4311 .set = gfx_v9_0_set_priv_inst_fault_state,
4312 .process = gfx_v9_0_priv_inst_irq,
4315 static void gfx_v9_0_set_irq_funcs(struct amdgpu_device *adev)
4317 adev->gfx.eop_irq.num_types = AMDGPU_CP_IRQ_LAST;
4318 adev->gfx.eop_irq.funcs = &gfx_v9_0_eop_irq_funcs;
4320 adev->gfx.priv_reg_irq.num_types = 1;
4321 adev->gfx.priv_reg_irq.funcs = &gfx_v9_0_priv_reg_irq_funcs;
4323 adev->gfx.priv_inst_irq.num_types = 1;
4324 adev->gfx.priv_inst_irq.funcs = &gfx_v9_0_priv_inst_irq_funcs;
4326 adev->gfx.kiq.irq.num_types = AMDGPU_CP_KIQ_IRQ_LAST;
4327 adev->gfx.kiq.irq.funcs = &gfx_v9_0_kiq_irq_funcs;
4330 static void gfx_v9_0_set_rlc_funcs(struct amdgpu_device *adev)
4332 switch (adev->asic_type) {
4335 adev->gfx.rlc.funcs = &gfx_v9_0_rlc_funcs;
4342 static void gfx_v9_0_set_gds_init(struct amdgpu_device *adev)
4344 /* init asci gds info */
4345 adev->gds.mem.total_size = RREG32_SOC15(GC, 0, mmGDS_VMID0_SIZE);
4346 adev->gds.gws.total_size = 64;
4347 adev->gds.oa.total_size = 16;
4349 if (adev->gds.mem.total_size == 64 * 1024) {
4350 adev->gds.mem.gfx_partition_size = 4096;
4351 adev->gds.mem.cs_partition_size = 4096;
4353 adev->gds.gws.gfx_partition_size = 4;
4354 adev->gds.gws.cs_partition_size = 4;
4356 adev->gds.oa.gfx_partition_size = 4;
4357 adev->gds.oa.cs_partition_size = 1;
4359 adev->gds.mem.gfx_partition_size = 1024;
4360 adev->gds.mem.cs_partition_size = 1024;
4362 adev->gds.gws.gfx_partition_size = 16;
4363 adev->gds.gws.cs_partition_size = 16;
4365 adev->gds.oa.gfx_partition_size = 4;
4366 adev->gds.oa.cs_partition_size = 4;
4370 static void gfx_v9_0_set_user_cu_inactive_bitmap(struct amdgpu_device *adev,
4378 data = bitmap << GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_CUS__SHIFT;
4379 data &= GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_CUS_MASK;
4381 WREG32_SOC15(GC, 0, mmGC_USER_SHADER_ARRAY_CONFIG, data);
4384 static u32 gfx_v9_0_get_cu_active_bitmap(struct amdgpu_device *adev)
4388 data = RREG32_SOC15(GC, 0, mmCC_GC_SHADER_ARRAY_CONFIG);
4389 data |= RREG32_SOC15(GC, 0, mmGC_USER_SHADER_ARRAY_CONFIG);
4391 data &= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_CUS_MASK;
4392 data >>= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_CUS__SHIFT;
4394 mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_cu_per_sh);
4396 return (~data) & mask;
4399 static int gfx_v9_0_get_cu_info(struct amdgpu_device *adev,
4400 struct amdgpu_cu_info *cu_info)
4402 int i, j, k, counter, active_cu_number = 0;
4403 u32 mask, bitmap, ao_bitmap, ao_cu_mask = 0;
4404 unsigned disable_masks[4 * 2];
4406 if (!adev || !cu_info)
4409 amdgpu_gfx_parse_disable_cu(disable_masks, 4, 2);
4411 mutex_lock(&adev->grbm_idx_mutex);
4412 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
4413 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
4417 gfx_v9_0_select_se_sh(adev, i, j, 0xffffffff);
4419 gfx_v9_0_set_user_cu_inactive_bitmap(
4420 adev, disable_masks[i * 2 + j]);
4421 bitmap = gfx_v9_0_get_cu_active_bitmap(adev);
4422 cu_info->bitmap[i][j] = bitmap;
4424 for (k = 0; k < adev->gfx.config.max_cu_per_sh; k ++) {
4425 if (bitmap & mask) {
4426 if (counter < adev->gfx.config.max_cu_per_sh)
4432 active_cu_number += counter;
4434 ao_cu_mask |= (ao_bitmap << (i * 16 + j * 8));
4435 cu_info->ao_cu_bitmap[i][j] = ao_bitmap;
4438 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
4439 mutex_unlock(&adev->grbm_idx_mutex);
4441 cu_info->number = active_cu_number;
4442 cu_info->ao_cu_mask = ao_cu_mask;
4447 const struct amdgpu_ip_block_version gfx_v9_0_ip_block =
4449 .type = AMD_IP_BLOCK_TYPE_GFX,
4453 .funcs = &gfx_v9_0_ip_funcs,