3 * (C) COPYRIGHT 2011-2015 ARM Limited. All rights reserved.
5 * This program is free software and is provided to you under the terms of the
6 * GNU General Public License version 2 as published by the Free Software
7 * Foundation, and any use by you of this program is subject to the terms
10 * A copy of the licence is included with the program, and can also be obtained
11 * from Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor,
12 * Boston, MA 02110-1301, USA.
21 * Base kernel property query APIs
24 #include <mali_kbase.h>
25 #include <mali_midg_regmap.h>
26 #include <mali_kbase_gpuprops.h>
27 #include <mali_kbase_config_defaults.h>
28 #include <mali_kbase_hwaccess_gpuprops.h>
29 #include <linux/clk.h>
32 * KBASE_UBFX32 - Extracts bits from a 32-bit bitfield.
33 * @value: The value from which to extract bits.
34 * @offset: The first bit to extract (0 being the LSB).
35 * @size: The number of bits to extract.
37 * Context: @offset + @size <= 32.
39 * Return: Bits [@offset, @offset + @size) from @value.
41 /* from mali_cdsb.h */
42 #define KBASE_UBFX32(value, offset, size) \
43 (((u32)(value) >> (u32)(offset)) & (u32)((1ULL << (u32)(size)) - 1))
45 int kbase_gpuprops_uk_get_props(struct kbase_context *kctx, struct kbase_uk_gpuprops * const kbase_props)
47 kbase_gpu_clk_speed_func get_gpu_speed_mhz;
51 KBASE_DEBUG_ASSERT(NULL != kctx);
52 KBASE_DEBUG_ASSERT(NULL != kbase_props);
54 /* Current GPU speed is requested from the system integrator via the GPU_SPEED_FUNC function.
55 * If that function fails, or the function is not provided by the system integrator, we report the maximum
56 * GPU speed as specified by GPU_FREQ_KHZ_MAX.
58 get_gpu_speed_mhz = (kbase_gpu_clk_speed_func) GPU_SPEED_FUNC;
59 if (get_gpu_speed_mhz != NULL) {
60 rc = get_gpu_speed_mhz(&gpu_speed_mhz);
61 #ifdef CONFIG_MALI_DEBUG
62 /* Issue a warning message when the reported GPU speed falls outside the min/max range */
64 u32 gpu_speed_khz = gpu_speed_mhz * 1000;
66 if (gpu_speed_khz < kctx->kbdev->gpu_props.props.core_props.gpu_freq_khz_min ||
67 gpu_speed_khz > kctx->kbdev->gpu_props.props.core_props.gpu_freq_khz_max)
68 dev_warn(kctx->kbdev->dev, "GPU Speed is outside of min/max range (got %lu Khz, min %lu Khz, max %lu Khz)\n",
69 (unsigned long)gpu_speed_khz,
70 (unsigned long)kctx->kbdev->gpu_props.props.core_props.gpu_freq_khz_min,
71 (unsigned long)kctx->kbdev->gpu_props.props.core_props.gpu_freq_khz_max);
73 #endif /* CONFIG_MALI_DEBUG */
75 if (kctx->kbdev->clock) {
76 gpu_speed_mhz = clk_get_rate(kctx->kbdev->clock) / 1000000;
80 gpu_speed_mhz = kctx->kbdev->gpu_props.props.core_props.gpu_freq_khz_max / 1000;
82 kctx->kbdev->gpu_props.props.core_props.gpu_speed_mhz = gpu_speed_mhz;
84 memcpy(&kbase_props->props, &kctx->kbdev->gpu_props.props, sizeof(kbase_props->props));
86 /* Before API 8.2 they expect L3 cache info here, which was always 0 */
87 if (kctx->api_version < KBASE_API_VERSION(8, 2))
88 kbase_props->props.raw_props.suspend_size = 0;
93 static void kbase_gpuprops_construct_coherent_groups(base_gpu_props * const props)
95 struct mali_base_gpu_coherent_group *current_group;
98 u64 first_set, first_set_prev;
101 KBASE_DEBUG_ASSERT(NULL != props);
103 props->coherency_info.coherency = props->raw_props.mem_features;
104 props->coherency_info.num_core_groups = hweight64(props->raw_props.l2_present);
106 if (props->coherency_info.coherency & GROUPS_L2_COHERENT) {
107 /* Group is l2 coherent */
108 group_present = props->raw_props.l2_present;
110 /* Group is l1 coherent */
111 group_present = props->raw_props.shader_present;
115 * The coherent group mask can be computed from the l2 present
118 * For the coherent group n:
119 * group_mask[n] = (first_set[n] - 1) & ~(first_set[n-1] - 1)
120 * where first_set is group_present with only its nth set-bit kept
121 * (i.e. the position from where a new group starts).
123 * For instance if the groups are l2 coherent and l2_present=0x0..01111:
125 * group_mask[1] = (first_set[1] - 1) & ~(first_set[0] - 1)
126 * = (0x0..010 - 1) & ~(0x0..01 - 1)
128 * The second mask is:
129 * group_mask[2] = (first_set[2] - 1) & ~(first_set[1] - 1)
130 * = (0x0..100 - 1) & ~(0x0..010 - 1)
132 * And so on until all the bits from group_present have been cleared
133 * (i.e. there is no group left).
136 current_group = props->coherency_info.group;
137 first_set = group_present & ~(group_present - 1);
139 while (group_present != 0 && num_groups < BASE_MAX_COHERENT_GROUPS) {
140 group_present -= first_set; /* Clear the current group bit */
141 first_set_prev = first_set;
143 first_set = group_present & ~(group_present - 1);
144 group_mask = (first_set - 1) & ~(first_set_prev - 1);
146 /* Populate the coherent_group structure for each group */
147 current_group->core_mask = group_mask & props->raw_props.shader_present;
148 current_group->num_cores = hweight64(current_group->core_mask);
154 if (group_present != 0)
155 pr_warn("Too many coherent groups (keeping only %d groups).\n", BASE_MAX_COHERENT_GROUPS);
157 props->coherency_info.num_groups = num_groups;
161 * kbase_gpuprops_get_props - Get the GPU configuration
162 * @gpu_props: The &base_gpu_props structure
163 * @kbdev: The &struct kbase_device structure for the device
165 * Fill the &base_gpu_props structure with values from the GPU configuration
166 * registers. Only the raw properties are filled in this function
168 static void kbase_gpuprops_get_props(base_gpu_props * const gpu_props, struct kbase_device *kbdev)
170 struct kbase_gpuprops_regdump regdump;
173 KBASE_DEBUG_ASSERT(NULL != kbdev);
174 KBASE_DEBUG_ASSERT(NULL != gpu_props);
176 /* Dump relevant registers */
177 kbase_backend_gpuprops_get(kbdev, ®dump);
179 gpu_props->raw_props.gpu_id = regdump.gpu_id;
180 gpu_props->raw_props.tiler_features = regdump.tiler_features;
181 gpu_props->raw_props.mem_features = regdump.mem_features;
182 gpu_props->raw_props.mmu_features = regdump.mmu_features;
183 gpu_props->raw_props.l2_features = regdump.l2_features;
184 gpu_props->raw_props.suspend_size = regdump.suspend_size;
186 gpu_props->raw_props.as_present = regdump.as_present;
187 gpu_props->raw_props.js_present = regdump.js_present;
188 gpu_props->raw_props.shader_present = ((u64) regdump.shader_present_hi << 32) + regdump.shader_present_lo;
189 gpu_props->raw_props.tiler_present = ((u64) regdump.tiler_present_hi << 32) + regdump.tiler_present_lo;
190 gpu_props->raw_props.l2_present = ((u64) regdump.l2_present_hi << 32) + regdump.l2_present_lo;
192 for (i = 0; i < GPU_MAX_JOB_SLOTS; i++)
193 gpu_props->raw_props.js_features[i] = regdump.js_features[i];
195 for (i = 0; i < BASE_GPU_NUM_TEXTURE_FEATURES_REGISTERS; i++)
196 gpu_props->raw_props.texture_features[i] = regdump.texture_features[i];
198 gpu_props->raw_props.thread_max_barrier_size = regdump.thread_max_barrier_size;
199 gpu_props->raw_props.thread_max_threads = regdump.thread_max_threads;
200 gpu_props->raw_props.thread_max_workgroup_size = regdump.thread_max_workgroup_size;
201 gpu_props->raw_props.thread_features = regdump.thread_features;
205 * kbase_gpuprops_calculate_props - Calculate the derived properties
206 * @gpu_props: The &base_gpu_props structure
207 * @kbdev: The &struct kbase_device structure for the device
209 * Fill the &base_gpu_props structure with values derived from the GPU
210 * configuration registers
212 static void kbase_gpuprops_calculate_props(base_gpu_props * const gpu_props, struct kbase_device *kbdev)
216 /* Populate the base_gpu_props structure */
217 gpu_props->core_props.version_status = KBASE_UBFX32(gpu_props->raw_props.gpu_id, 0U, 4);
218 gpu_props->core_props.minor_revision = KBASE_UBFX32(gpu_props->raw_props.gpu_id, 4U, 8);
219 gpu_props->core_props.major_revision = KBASE_UBFX32(gpu_props->raw_props.gpu_id, 12U, 4);
220 gpu_props->core_props.product_id = KBASE_UBFX32(gpu_props->raw_props.gpu_id, 16U, 16);
221 gpu_props->core_props.log2_program_counter_size = KBASE_GPU_PC_SIZE_LOG2;
222 gpu_props->core_props.gpu_available_memory_size = totalram_pages << PAGE_SHIFT;
224 for (i = 0; i < BASE_GPU_NUM_TEXTURE_FEATURES_REGISTERS; i++)
225 gpu_props->core_props.texture_features[i] = gpu_props->raw_props.texture_features[i];
227 gpu_props->l2_props.log2_line_size = KBASE_UBFX32(gpu_props->raw_props.l2_features, 0U, 8);
228 gpu_props->l2_props.log2_cache_size = KBASE_UBFX32(gpu_props->raw_props.l2_features, 16U, 8);
230 /* Field with number of l2 slices is added to MEM_FEATURES register
231 * since t76x. Below code assumes that for older GPU reserved bits will
232 * be read as zero. */
233 gpu_props->l2_props.num_l2_slices =
234 KBASE_UBFX32(gpu_props->raw_props.mem_features, 8U, 4) + 1;
236 gpu_props->tiler_props.bin_size_bytes = 1 << KBASE_UBFX32(gpu_props->raw_props.tiler_features, 0U, 6);
237 gpu_props->tiler_props.max_active_levels = KBASE_UBFX32(gpu_props->raw_props.tiler_features, 8U, 4);
239 if (gpu_props->raw_props.thread_max_threads == 0)
240 gpu_props->thread_props.max_threads = THREAD_MT_DEFAULT;
242 gpu_props->thread_props.max_threads = gpu_props->raw_props.thread_max_threads;
244 if (gpu_props->raw_props.thread_max_workgroup_size == 0)
245 gpu_props->thread_props.max_workgroup_size = THREAD_MWS_DEFAULT;
247 gpu_props->thread_props.max_workgroup_size = gpu_props->raw_props.thread_max_workgroup_size;
249 if (gpu_props->raw_props.thread_max_barrier_size == 0)
250 gpu_props->thread_props.max_barrier_size = THREAD_MBS_DEFAULT;
252 gpu_props->thread_props.max_barrier_size = gpu_props->raw_props.thread_max_barrier_size;
254 gpu_props->thread_props.max_registers = KBASE_UBFX32(gpu_props->raw_props.thread_features, 0U, 16);
255 gpu_props->thread_props.max_task_queue = KBASE_UBFX32(gpu_props->raw_props.thread_features, 16U, 8);
256 gpu_props->thread_props.max_thread_group_split = KBASE_UBFX32(gpu_props->raw_props.thread_features, 24U, 6);
257 gpu_props->thread_props.impl_tech = KBASE_UBFX32(gpu_props->raw_props.thread_features, 30U, 2);
259 /* If values are not specified, then use defaults */
260 if (gpu_props->thread_props.max_registers == 0) {
261 gpu_props->thread_props.max_registers = THREAD_MR_DEFAULT;
262 gpu_props->thread_props.max_task_queue = THREAD_MTQ_DEFAULT;
263 gpu_props->thread_props.max_thread_group_split = THREAD_MTGS_DEFAULT;
265 /* Initialize the coherent_group structure for each group */
266 kbase_gpuprops_construct_coherent_groups(gpu_props);
269 void kbase_gpuprops_set(struct kbase_device *kbdev)
271 struct kbase_gpu_props *gpu_props;
272 struct gpu_raw_gpu_props *raw;
274 KBASE_DEBUG_ASSERT(NULL != kbdev);
275 gpu_props = &kbdev->gpu_props;
276 raw = &gpu_props->props.raw_props;
278 /* Initialize the base_gpu_props structure from the hardware */
279 kbase_gpuprops_get_props(&gpu_props->props, kbdev);
281 /* Populate the derived properties */
282 kbase_gpuprops_calculate_props(&gpu_props->props, kbdev);
284 /* Populate kbase-only fields */
285 gpu_props->l2_props.associativity = KBASE_UBFX32(raw->l2_features, 8U, 8);
286 gpu_props->l2_props.external_bus_width = KBASE_UBFX32(raw->l2_features, 24U, 8);
288 gpu_props->mem.core_group = KBASE_UBFX32(raw->mem_features, 0U, 1);
290 gpu_props->mmu.va_bits = KBASE_UBFX32(raw->mmu_features, 0U, 8);
291 gpu_props->mmu.pa_bits = KBASE_UBFX32(raw->mmu_features, 8U, 8);
293 gpu_props->num_cores = hweight64(raw->shader_present);
294 gpu_props->num_core_groups = hweight64(raw->l2_present);
295 gpu_props->num_address_spaces = hweight32(raw->as_present);
296 gpu_props->num_job_slots = hweight32(raw->js_present);
299 void kbase_gpuprops_set_features(struct kbase_device *kbdev)
301 base_gpu_props *gpu_props;
302 struct kbase_gpuprops_regdump regdump;
304 gpu_props = &kbdev->gpu_props.props;
306 /* Dump relevant registers */
307 kbase_backend_gpuprops_get_features(kbdev, ®dump);
310 * Copy the raw value from the register, later this will get turned
311 * into the selected coherency mode.
313 gpu_props->raw_props.coherency_mode = regdump.coherency_features;