1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include "gen7_renderclear.h"
7 #include "i915_drv.h"
8 #include "intel_gpu_commands.h"
9 
10 #define MAX_URB_ENTRIES 64
11 #define STATE_SIZE (4 * 1024)
12 #define GT3_INLINE_DATA_DELAYS 0x1E00
13 #define batch_advance(Y, CS) GEM_BUG_ON((Y)->end != (CS))
14 
15 struct cb_kernel {
16 	const void *data;
17 	u32 size;
18 };
19 
20 #define CB_KERNEL(name) { .data = (name), .size = sizeof(name) }
21 
22 #include "ivb_clear_kernel.c"
23 static const struct cb_kernel cb_kernel_ivb = CB_KERNEL(ivb_clear_kernel);
24 
25 #include "hsw_clear_kernel.c"
26 static const struct cb_kernel cb_kernel_hsw = CB_KERNEL(hsw_clear_kernel);
27 
28 struct batch_chunk {
29 	struct i915_vma *vma;
30 	u32 offset;
31 	u32 *start;
32 	u32 *end;
33 	u32 max_items;
34 };
35 
36 struct batch_vals {
37 	u32 max_primitives;
38 	u32 max_urb_entries;
39 	u32 cmd_size;
40 	u32 state_size;
41 	u32 state_start;
42 	u32 batch_size;
43 	u32 surface_height;
44 	u32 surface_width;
45 	u32 scratch_size;
46 	u32 max_size;
47 };
48 
49 static void
50 batch_get_defaults(struct drm_i915_private *i915, struct batch_vals *bv)
51 {
52 	if (IS_HASWELL(i915)) {
53 		bv->max_primitives = 280;
54 		bv->max_urb_entries = MAX_URB_ENTRIES;
55 		bv->surface_height = 16 * 16;
56 		bv->surface_width = 32 * 2 * 16;
57 	} else {
58 		bv->max_primitives = 128;
59 		bv->max_urb_entries = MAX_URB_ENTRIES / 2;
60 		bv->surface_height = 16 * 8;
61 		bv->surface_width = 32 * 16;
62 	}
63 	bv->cmd_size = bv->max_primitives * 4096;
64 	bv->state_size = STATE_SIZE;
65 	bv->state_start = bv->cmd_size;
66 	bv->batch_size = bv->cmd_size + bv->state_size;
67 	bv->scratch_size = bv->surface_height * bv->surface_width;
68 	bv->max_size = bv->batch_size + bv->scratch_size;
69 }
70 
71 static void batch_init(struct batch_chunk *bc,
72 		       struct i915_vma *vma,
73 		       u32 *start, u32 offset, u32 max_bytes)
74 {
75 	bc->vma = vma;
76 	bc->offset = offset;
77 	bc->start = start + bc->offset / sizeof(*bc->start);
78 	bc->end = bc->start;
79 	bc->max_items = max_bytes / sizeof(*bc->start);
80 }
81 
82 static u32 batch_offset(const struct batch_chunk *bc, u32 *cs)
83 {
84 	return (cs - bc->start) * sizeof(*bc->start) + bc->offset;
85 }
86 
87 static u32 batch_addr(const struct batch_chunk *bc)
88 {
89 	return bc->vma->node.start;
90 }
91 
92 static void batch_add(struct batch_chunk *bc, const u32 d)
93 {
94 	GEM_BUG_ON((bc->end - bc->start) >= bc->max_items);
95 	*bc->end++ = d;
96 }
97 
98 static u32 *batch_alloc_items(struct batch_chunk *bc, u32 align, u32 items)
99 {
100 	u32 *map;
101 
102 	if (align) {
103 		u32 *end = PTR_ALIGN(bc->end, align);
104 
105 		memset32(bc->end, 0, end - bc->end);
106 		bc->end = end;
107 	}
108 
109 	map = bc->end;
110 	bc->end += items;
111 
112 	return map;
113 }
114 
115 static u32 *batch_alloc_bytes(struct batch_chunk *bc, u32 align, u32 bytes)
116 {
117 	GEM_BUG_ON(!IS_ALIGNED(bytes, sizeof(*bc->start)));
118 	return batch_alloc_items(bc, align, bytes / sizeof(*bc->start));
119 }
120 
121 static u32
122 gen7_fill_surface_state(struct batch_chunk *state,
123 			const u32 dst_offset,
124 			const struct batch_vals *bv)
125 {
126 	u32 surface_h = bv->surface_height;
127 	u32 surface_w = bv->surface_width;
128 	u32 *cs = batch_alloc_items(state, 32, 8);
129 	u32 offset = batch_offset(state, cs);
130 
131 #define SURFACE_2D 1
132 #define SURFACEFORMAT_B8G8R8A8_UNORM 0x0C0
133 #define RENDER_CACHE_READ_WRITE 1
134 
135 	*cs++ = SURFACE_2D << 29 |
136 		(SURFACEFORMAT_B8G8R8A8_UNORM << 18) |
137 		(RENDER_CACHE_READ_WRITE << 8);
138 
139 	*cs++ = batch_addr(state) + dst_offset;
140 
141 	*cs++ = ((surface_h / 4 - 1) << 16) | (surface_w / 4 - 1);
142 	*cs++ = surface_w;
143 	*cs++ = 0;
144 	*cs++ = 0;
145 	*cs++ = 0;
146 #define SHADER_CHANNELS(r, g, b, a) \
147 	(((r) << 25) | ((g) << 22) | ((b) << 19) | ((a) << 16))
148 	*cs++ = SHADER_CHANNELS(4, 5, 6, 7);
149 	batch_advance(state, cs);
150 
151 	return offset;
152 }
153 
154 static u32
155 gen7_fill_binding_table(struct batch_chunk *state,
156 			const struct batch_vals *bv)
157 {
158 	u32 surface_start = gen7_fill_surface_state(state, bv->batch_size, bv);
159 	u32 *cs = batch_alloc_items(state, 32, 8);
160 	u32 offset = batch_offset(state, cs);
161 
162 	*cs++ = surface_start - state->offset;
163 	*cs++ = 0;
164 	*cs++ = 0;
165 	*cs++ = 0;
166 	*cs++ = 0;
167 	*cs++ = 0;
168 	*cs++ = 0;
169 	*cs++ = 0;
170 	batch_advance(state, cs);
171 
172 	return offset;
173 }
174 
175 static u32
176 gen7_fill_kernel_data(struct batch_chunk *state,
177 		      const u32 *data,
178 		      const u32 size)
179 {
180 	return batch_offset(state,
181 			    memcpy(batch_alloc_bytes(state, 64, size),
182 				   data, size));
183 }
184 
185 static u32
186 gen7_fill_interface_descriptor(struct batch_chunk *state,
187 			       const struct batch_vals *bv,
188 			       const struct cb_kernel *kernel,
189 			       unsigned int count)
190 {
191 	u32 kernel_offset =
192 		gen7_fill_kernel_data(state, kernel->data, kernel->size);
193 	u32 binding_table = gen7_fill_binding_table(state, bv);
194 	u32 *cs = batch_alloc_items(state, 32, 8 * count);
195 	u32 offset = batch_offset(state, cs);
196 
197 	*cs++ = kernel_offset;
198 	*cs++ = (1 << 7) | (1 << 13);
199 	*cs++ = 0;
200 	*cs++ = (binding_table - state->offset) | 1;
201 	*cs++ = 0;
202 	*cs++ = 0;
203 	*cs++ = 0;
204 	*cs++ = 0;
205 
206 	/* 1 - 63dummy idds */
207 	memset32(cs, 0x00, (count - 1) * 8);
208 	batch_advance(state, cs + (count - 1) * 8);
209 
210 	return offset;
211 }
212 
213 static void
214 gen7_emit_state_base_address(struct batch_chunk *batch,
215 			     u32 surface_state_base)
216 {
217 	u32 *cs = batch_alloc_items(batch, 0, 12);
218 
219 	*cs++ = STATE_BASE_ADDRESS | (12 - 2);
220 	/* general */
221 	*cs++ = batch_addr(batch) | BASE_ADDRESS_MODIFY;
222 	/* surface */
223 	*cs++ = batch_addr(batch) | surface_state_base | BASE_ADDRESS_MODIFY;
224 	/* dynamic */
225 	*cs++ = batch_addr(batch) | BASE_ADDRESS_MODIFY;
226 	/* indirect */
227 	*cs++ = batch_addr(batch) | BASE_ADDRESS_MODIFY;
228 	/* instruction */
229 	*cs++ = batch_addr(batch) | BASE_ADDRESS_MODIFY;
230 
231 	/* general/dynamic/indirect/instruction access Bound */
232 	*cs++ = 0;
233 	*cs++ = BASE_ADDRESS_MODIFY;
234 	*cs++ = 0;
235 	*cs++ = BASE_ADDRESS_MODIFY;
236 	*cs++ = 0;
237 	*cs++ = 0;
238 	batch_advance(batch, cs);
239 }
240 
241 static void
242 gen7_emit_vfe_state(struct batch_chunk *batch,
243 		    const struct batch_vals *bv,
244 		    u32 urb_size, u32 curbe_size,
245 		    u32 mode)
246 {
247 	u32 urb_entries = bv->max_urb_entries;
248 	u32 threads = bv->max_primitives - 1;
249 	u32 *cs = batch_alloc_items(batch, 32, 8);
250 
251 	*cs++ = MEDIA_VFE_STATE | (8 - 2);
252 
253 	/* scratch buffer */
254 	*cs++ = 0;
255 
256 	/* number of threads & urb entries for GPGPU vs Media Mode */
257 	*cs++ = threads << 16 | urb_entries << 8 | mode << 2;
258 
259 	*cs++ = 0;
260 
261 	/* urb entry size & curbe size in 256 bits unit */
262 	*cs++ = urb_size << 16 | curbe_size;
263 
264 	/* scoreboard */
265 	*cs++ = 0;
266 	*cs++ = 0;
267 	*cs++ = 0;
268 	batch_advance(batch, cs);
269 }
270 
271 static void
272 gen7_emit_interface_descriptor_load(struct batch_chunk *batch,
273 				    const u32 interface_descriptor,
274 				    unsigned int count)
275 {
276 	u32 *cs = batch_alloc_items(batch, 8, 4);
277 
278 	*cs++ = MEDIA_INTERFACE_DESCRIPTOR_LOAD | (4 - 2);
279 	*cs++ = 0;
280 	*cs++ = count * 8 * sizeof(*cs);
281 
282 	/*
283 	 * interface descriptor address - it is relative to the dynamics base
284 	 * address
285 	 */
286 	*cs++ = interface_descriptor;
287 	batch_advance(batch, cs);
288 }
289 
290 static void
291 gen7_emit_media_object(struct batch_chunk *batch,
292 		       unsigned int media_object_index)
293 {
294 	unsigned int x_offset = (media_object_index % 16) * 64;
295 	unsigned int y_offset = (media_object_index / 16) * 16;
296 	unsigned int inline_data_size;
297 	unsigned int media_batch_size;
298 	unsigned int i;
299 	u32 *cs;
300 
301 	inline_data_size = 112 * 8;
302 	media_batch_size = inline_data_size + 6;
303 
304 	cs = batch_alloc_items(batch, 8, media_batch_size);
305 
306 	*cs++ = MEDIA_OBJECT | (media_batch_size - 2);
307 
308 	/* interface descriptor offset */
309 	*cs++ = 0;
310 
311 	/* without indirect data */
312 	*cs++ = 0;
313 	*cs++ = 0;
314 
315 	/* scoreboard */
316 	*cs++ = 0;
317 	*cs++ = 0;
318 
319 	/* inline */
320 	*cs++ = (y_offset << 16) | (x_offset);
321 	*cs++ = 0;
322 	*cs++ = GT3_INLINE_DATA_DELAYS;
323 	for (i = 3; i < inline_data_size; i++)
324 		*cs++ = 0;
325 
326 	batch_advance(batch, cs);
327 }
328 
329 static void gen7_emit_pipeline_flush(struct batch_chunk *batch)
330 {
331 	u32 *cs = batch_alloc_items(batch, 0, 5);
332 
333 	*cs++ = GFX_OP_PIPE_CONTROL(5);
334 	*cs++ = PIPE_CONTROL_STATE_CACHE_INVALIDATE |
335 		PIPE_CONTROL_GLOBAL_GTT_IVB;
336 	*cs++ = 0;
337 	*cs++ = 0;
338 	*cs++ = 0;
339 	batch_advance(batch, cs);
340 }
341 
342 static void emit_batch(struct i915_vma * const vma,
343 		       u32 *start,
344 		       const struct batch_vals *bv)
345 {
346 	struct drm_i915_private *i915 = vma->vm->i915;
347 	unsigned int desc_count = 64;
348 	const u32 urb_size = 112;
349 	struct batch_chunk cmds, state;
350 	u32 interface_descriptor;
351 	unsigned int i;
352 
353 	batch_init(&cmds, vma, start, 0, bv->cmd_size);
354 	batch_init(&state, vma, start, bv->state_start, bv->state_size);
355 
356 	interface_descriptor =
357 		gen7_fill_interface_descriptor(&state, bv,
358 					       IS_HASWELL(i915) ?
359 					       &cb_kernel_hsw :
360 					       &cb_kernel_ivb,
361 					       desc_count);
362 	gen7_emit_pipeline_flush(&cmds);
363 	batch_add(&cmds, PIPELINE_SELECT | PIPELINE_SELECT_MEDIA);
364 	batch_add(&cmds, MI_NOOP);
365 	gen7_emit_state_base_address(&cmds, interface_descriptor);
366 	gen7_emit_pipeline_flush(&cmds);
367 
368 	gen7_emit_vfe_state(&cmds, bv, urb_size - 1, 0, 0);
369 
370 	gen7_emit_interface_descriptor_load(&cmds,
371 					    interface_descriptor,
372 					    desc_count);
373 
374 	for (i = 0; i < bv->max_primitives; i++)
375 		gen7_emit_media_object(&cmds, i);
376 
377 	batch_add(&cmds, MI_BATCH_BUFFER_END);
378 }
379 
380 int gen7_setup_clear_gpr_bb(struct intel_engine_cs * const engine,
381 			    struct i915_vma * const vma)
382 {
383 	struct batch_vals bv;
384 	u32 *batch;
385 
386 	batch_get_defaults(engine->i915, &bv);
387 	if (!vma)
388 		return bv.max_size;
389 
390 	GEM_BUG_ON(vma->obj->base.size < bv.max_size);
391 
392 	batch = i915_gem_object_pin_map(vma->obj, I915_MAP_WC);
393 	if (IS_ERR(batch))
394 		return PTR_ERR(batch);
395 
396 	emit_batch(vma, memset(batch, 0, bv.max_size), &bv);
397 
398 	i915_gem_object_flush_map(vma->obj);
399 	__i915_gem_object_release_map(vma->obj);
400 
401 	return 0;
402 }
403