• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3  * Copyright 2014,2015 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * on the rights to use, copy, modify, merge, publish, distribute, sub
9  * license, and/or sell copies of the Software, and to permit persons to whom
10  * the Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors:
25  *      Jerome Glisse
26  */
27 
28 #include "sid.h"
29 #include "si_pipe.h"
30 
cik_sdma_copy_buffer(struct si_context * ctx,struct pipe_resource * dst,struct pipe_resource * src,uint64_t dst_offset,uint64_t src_offset,uint64_t size)31 static void cik_sdma_copy_buffer(struct si_context *ctx,
32 				 struct pipe_resource *dst,
33 				 struct pipe_resource *src,
34 				 uint64_t dst_offset,
35 				 uint64_t src_offset,
36 				 uint64_t size)
37 {
38 	struct radeon_winsys_cs *cs = ctx->b.dma.cs;
39 	unsigned i, ncopy, csize;
40 	struct r600_resource *rdst = r600_resource(dst);
41 	struct r600_resource *rsrc = r600_resource(src);
42 
43 	/* Mark the buffer range of destination as valid (initialized),
44 	 * so that transfer_map knows it should wait for the GPU when mapping
45 	 * that range. */
46 	util_range_add(&rdst->valid_buffer_range, dst_offset,
47 		       dst_offset + size);
48 
49 	dst_offset += rdst->gpu_address;
50 	src_offset += rsrc->gpu_address;
51 
52 	ncopy = DIV_ROUND_UP(size, CIK_SDMA_COPY_MAX_SIZE);
53 	r600_need_dma_space(&ctx->b, ncopy * 7, rdst, rsrc);
54 
55 	for (i = 0; i < ncopy; i++) {
56 		csize = MIN2(size, CIK_SDMA_COPY_MAX_SIZE);
57 		radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
58 						CIK_SDMA_COPY_SUB_OPCODE_LINEAR,
59 						0));
60 		radeon_emit(cs, csize);
61 		radeon_emit(cs, 0); /* src/dst endian swap */
62 		radeon_emit(cs, src_offset);
63 		radeon_emit(cs, src_offset >> 32);
64 		radeon_emit(cs, dst_offset);
65 		radeon_emit(cs, dst_offset >> 32);
66 		dst_offset += csize;
67 		src_offset += csize;
68 		size -= csize;
69 	}
70 }
71 
cik_sdma_clear_buffer(struct pipe_context * ctx,struct pipe_resource * dst,uint64_t offset,uint64_t size,unsigned clear_value)72 static void cik_sdma_clear_buffer(struct pipe_context *ctx,
73 				  struct pipe_resource *dst,
74 				  uint64_t offset,
75 				  uint64_t size,
76 				  unsigned clear_value)
77 {
78 	struct si_context *sctx = (struct si_context *)ctx;
79 	struct radeon_winsys_cs *cs = sctx->b.dma.cs;
80 	unsigned i, ncopy, csize;
81 	struct r600_resource *rdst = r600_resource(dst);
82 
83 	if (!cs || offset % 4 != 0 || size % 4 != 0) {
84 		ctx->clear_buffer(ctx, dst, offset, size, &clear_value, 4);
85 		return;
86 	}
87 
88 	/* Mark the buffer range of destination as valid (initialized),
89 	 * so that transfer_map knows it should wait for the GPU when mapping
90 	 * that range. */
91 	util_range_add(&rdst->valid_buffer_range, offset, offset + size);
92 
93 	offset += rdst->gpu_address;
94 
95 	/* the same maximum size as for copying */
96 	ncopy = DIV_ROUND_UP(size, CIK_SDMA_COPY_MAX_SIZE);
97 	r600_need_dma_space(&sctx->b, ncopy * 5, rdst, NULL);
98 
99 	for (i = 0; i < ncopy; i++) {
100 		csize = MIN2(size, CIK_SDMA_COPY_MAX_SIZE);
101 		radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_PACKET_CONSTANT_FILL, 0,
102 						0x8000 /* dword copy */));
103 		radeon_emit(cs, offset);
104 		radeon_emit(cs, offset >> 32);
105 		radeon_emit(cs, clear_value);
106 		radeon_emit(cs, csize);
107 		offset += csize;
108 		size -= csize;
109 	}
110 }
111 
minify_as_blocks(unsigned width,unsigned level,unsigned blk_w)112 static unsigned minify_as_blocks(unsigned width, unsigned level, unsigned blk_w)
113 {
114 	width = u_minify(width, level);
115 	return DIV_ROUND_UP(width, blk_w);
116 }
117 
encode_tile_info(struct si_context * sctx,struct r600_texture * tex,unsigned level,bool set_bpp)118 static unsigned encode_tile_info(struct si_context *sctx,
119 				 struct r600_texture *tex, unsigned level,
120 				 bool set_bpp)
121 {
122 	struct radeon_info *info = &sctx->screen->b.info;
123 	unsigned tile_index = tex->surface.tiling_index[level];
124 	unsigned macro_tile_index = tex->surface.macro_tile_index;
125 	unsigned tile_mode = info->si_tile_mode_array[tile_index];
126 	unsigned macro_tile_mode = info->cik_macrotile_mode_array[macro_tile_index];
127 
128 	return (set_bpp ? util_logbase2(tex->surface.bpe) : 0) |
129 		(G_009910_ARRAY_MODE(tile_mode) << 3) |
130 		(G_009910_MICRO_TILE_MODE_NEW(tile_mode) << 8) |
131 		/* Non-depth modes don't have TILE_SPLIT set. */
132 		((util_logbase2(tex->surface.tile_split >> 6)) << 11) |
133 		(G_009990_BANK_WIDTH(macro_tile_mode) << 15) |
134 		(G_009990_BANK_HEIGHT(macro_tile_mode) << 18) |
135 		(G_009990_NUM_BANKS(macro_tile_mode) << 21) |
136 		(G_009990_MACRO_TILE_ASPECT(macro_tile_mode) << 24) |
137 		(G_009910_PIPE_CONFIG(tile_mode) << 26);
138 }
139 
cik_sdma_copy_texture(struct si_context * sctx,struct pipe_resource * dst,unsigned dst_level,unsigned dstx,unsigned dsty,unsigned dstz,struct pipe_resource * src,unsigned src_level,const struct pipe_box * src_box)140 static bool cik_sdma_copy_texture(struct si_context *sctx,
141 				  struct pipe_resource *dst,
142 				  unsigned dst_level,
143 				  unsigned dstx, unsigned dsty, unsigned dstz,
144 				  struct pipe_resource *src,
145 				  unsigned src_level,
146 				  const struct pipe_box *src_box)
147 {
148 	struct radeon_info *info = &sctx->screen->b.info;
149 	struct r600_texture *rsrc = (struct r600_texture*)src;
150 	struct r600_texture *rdst = (struct r600_texture*)dst;
151 	unsigned bpp = rdst->surface.bpe;
152 	uint64_t dst_address = rdst->resource.gpu_address +
153 			       rdst->surface.level[dst_level].offset;
154 	uint64_t src_address = rsrc->resource.gpu_address +
155 			       rsrc->surface.level[src_level].offset;
156 	unsigned dst_mode = rdst->surface.level[dst_level].mode;
157 	unsigned src_mode = rsrc->surface.level[src_level].mode;
158 	unsigned dst_tile_index = rdst->surface.tiling_index[dst_level];
159 	unsigned src_tile_index = rsrc->surface.tiling_index[src_level];
160 	unsigned dst_tile_mode = info->si_tile_mode_array[dst_tile_index];
161 	unsigned src_tile_mode = info->si_tile_mode_array[src_tile_index];
162 	unsigned dst_micro_mode = G_009910_MICRO_TILE_MODE_NEW(dst_tile_mode);
163 	unsigned src_micro_mode = G_009910_MICRO_TILE_MODE_NEW(src_tile_mode);
164 	unsigned dst_pitch = rdst->surface.level[dst_level].nblk_x;
165 	unsigned src_pitch = rsrc->surface.level[src_level].nblk_x;
166 	uint64_t dst_slice_pitch = rdst->surface.level[dst_level].slice_size / bpp;
167 	uint64_t src_slice_pitch = rsrc->surface.level[src_level].slice_size / bpp;
168 	unsigned dst_width = minify_as_blocks(rdst->resource.b.b.width0,
169 					      dst_level, rdst->surface.blk_w);
170 	unsigned src_width = minify_as_blocks(rsrc->resource.b.b.width0,
171 					      src_level, rsrc->surface.blk_w);
172 	unsigned dst_height = minify_as_blocks(rdst->resource.b.b.height0,
173 					       dst_level, rdst->surface.blk_h);
174 	unsigned src_height = minify_as_blocks(rsrc->resource.b.b.height0,
175 					       src_level, rsrc->surface.blk_h);
176 	unsigned srcx = src_box->x / rsrc->surface.blk_w;
177 	unsigned srcy = src_box->y / rsrc->surface.blk_h;
178 	unsigned srcz = src_box->z;
179 	unsigned copy_width = DIV_ROUND_UP(src_box->width, rsrc->surface.blk_w);
180 	unsigned copy_height = DIV_ROUND_UP(src_box->height, rsrc->surface.blk_h);
181 	unsigned copy_depth = src_box->depth;
182 
183 	assert(src_level <= src->last_level);
184 	assert(dst_level <= dst->last_level);
185 	assert(rdst->surface.level[dst_level].offset +
186 	       dst_slice_pitch * bpp * (dstz + src_box->depth) <=
187 	       rdst->resource.buf->size);
188 	assert(rsrc->surface.level[src_level].offset +
189 	       src_slice_pitch * bpp * (srcz + src_box->depth) <=
190 	       rsrc->resource.buf->size);
191 
192 	if (!r600_prepare_for_dma_blit(&sctx->b, rdst, dst_level, dstx, dsty,
193 					dstz, rsrc, src_level, src_box))
194 		return false;
195 
196 	dstx /= rdst->surface.blk_w;
197 	dsty /= rdst->surface.blk_h;
198 
199 	if (srcx >= (1 << 14) ||
200 	    srcy >= (1 << 14) ||
201 	    srcz >= (1 << 11) ||
202 	    dstx >= (1 << 14) ||
203 	    dsty >= (1 << 14) ||
204 	    dstz >= (1 << 11))
205 		return false;
206 
207 	/* Linear -> linear sub-window copy. */
208 	if (dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED &&
209 	    src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED &&
210 	    /* check if everything fits into the bitfields */
211 	    src_pitch <= (1 << 14) &&
212 	    dst_pitch <= (1 << 14) &&
213 	    src_slice_pitch <= (1 << 28) &&
214 	    dst_slice_pitch <= (1 << 28) &&
215 	    copy_width <= (1 << 14) &&
216 	    copy_height <= (1 << 14) &&
217 	    copy_depth <= (1 << 11) &&
218 	    /* HW limitation - CIK: */
219 	    (sctx->b.chip_class != CIK ||
220 	     (copy_width < (1 << 14) &&
221 	      copy_height < (1 << 14) &&
222 	      copy_depth < (1 << 11))) &&
223 	    /* HW limitation - some CIK parts: */
224 	    ((sctx->b.family != CHIP_BONAIRE &&
225 	      sctx->b.family != CHIP_KAVERI) ||
226 	     (srcx + copy_width != (1 << 14) &&
227 	      srcy + copy_height != (1 << 14)))) {
228 		struct radeon_winsys_cs *cs = sctx->b.dma.cs;
229 
230 		r600_need_dma_space(&sctx->b, 13, &rdst->resource, &rsrc->resource);
231 
232 		radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
233 						CIK_SDMA_COPY_SUB_OPCODE_LINEAR_SUB_WINDOW, 0) |
234 			    (util_logbase2(bpp) << 29));
235 		radeon_emit(cs, src_address);
236 		radeon_emit(cs, src_address >> 32);
237 		radeon_emit(cs, srcx | (srcy << 16));
238 		radeon_emit(cs, srcz | ((src_pitch - 1) << 16));
239 		radeon_emit(cs, src_slice_pitch - 1);
240 		radeon_emit(cs, dst_address);
241 		radeon_emit(cs, dst_address >> 32);
242 		radeon_emit(cs, dstx | (dsty << 16));
243 		radeon_emit(cs, dstz | ((dst_pitch - 1) << 16));
244 		radeon_emit(cs, dst_slice_pitch - 1);
245 		if (sctx->b.chip_class == CIK) {
246 			radeon_emit(cs, copy_width | (copy_height << 16));
247 			radeon_emit(cs, copy_depth);
248 		} else {
249 			radeon_emit(cs, (copy_width - 1) | ((copy_height - 1) << 16));
250 			radeon_emit(cs, (copy_depth - 1));
251 		}
252 		return true;
253 	}
254 
255 	/* Tiled <-> linear sub-window copy. */
256 	if ((src_mode >= RADEON_SURF_MODE_1D) != (dst_mode >= RADEON_SURF_MODE_1D)) {
257 		struct r600_texture *tiled = src_mode >= RADEON_SURF_MODE_1D ? rsrc : rdst;
258 		struct r600_texture *linear = tiled == rsrc ? rdst : rsrc;
259 		unsigned tiled_level =	tiled	== rsrc ? src_level : dst_level;
260 		unsigned linear_level =	linear	== rsrc ? src_level : dst_level;
261 		unsigned tiled_x =	tiled	== rsrc ? srcx : dstx;
262 		unsigned linear_x =	linear  == rsrc ? srcx : dstx;
263 		unsigned tiled_y =	tiled	== rsrc ? srcy : dsty;
264 		unsigned linear_y =	linear  == rsrc ? srcy : dsty;
265 		unsigned tiled_z =	tiled	== rsrc ? srcz : dstz;
266 		unsigned linear_z =	linear  == rsrc ? srcz : dstz;
267 		unsigned tiled_width =	tiled	== rsrc ? src_width : dst_width;
268 		unsigned linear_width =	linear	== rsrc ? src_width : dst_width;
269 		unsigned tiled_pitch =	tiled	== rsrc ? src_pitch : dst_pitch;
270 		unsigned linear_pitch =	linear	== rsrc ? src_pitch : dst_pitch;
271 		unsigned tiled_slice_pitch  = tiled  == rsrc ? src_slice_pitch : dst_slice_pitch;
272 		unsigned linear_slice_pitch = linear == rsrc ? src_slice_pitch : dst_slice_pitch;
273 		uint64_t tiled_address =  tiled  == rsrc ? src_address : dst_address;
274 		uint64_t linear_address = linear == rsrc ? src_address : dst_address;
275 		unsigned tiled_micro_mode = tiled == rsrc ? src_micro_mode : dst_micro_mode;
276 
277 		assert(tiled_pitch % 8 == 0);
278 		assert(tiled_slice_pitch % 64 == 0);
279 		unsigned pitch_tile_max = tiled_pitch / 8 - 1;
280 		unsigned slice_tile_max = tiled_slice_pitch / 64 - 1;
281 		unsigned xalign = MAX2(1, 4 / bpp);
282 		unsigned copy_width_aligned = copy_width;
283 
284 		/* If the region ends at the last pixel and is unaligned, we
285 		 * can copy the remainder of the line that is not visible to
286 		 * make it aligned.
287 		 */
288 		if (copy_width % xalign != 0 &&
289 		    linear_x + copy_width == linear_width &&
290 		    tiled_x  + copy_width == tiled_width &&
291 		    linear_x + align(copy_width, xalign) <= linear_pitch &&
292 		    tiled_x  + align(copy_width, xalign) <= tiled_pitch)
293 			copy_width_aligned = align(copy_width, xalign);
294 
295 		/* HW limitations. */
296 		if ((sctx->b.family == CHIP_BONAIRE ||
297 		     sctx->b.family == CHIP_KAVERI) &&
298 		    linear_pitch - 1 == 0x3fff &&
299 		    bpp == 16)
300 			return false;
301 
302 		if (sctx->b.chip_class == CIK &&
303 		    (copy_width_aligned == (1 << 14) ||
304 		     copy_height == (1 << 14) ||
305 		     copy_depth == (1 << 11)))
306 			return false;
307 
308 		if ((sctx->b.family == CHIP_BONAIRE ||
309 		     sctx->b.family == CHIP_KAVERI ||
310 		     sctx->b.family == CHIP_KABINI ||
311 		     sctx->b.family == CHIP_MULLINS) &&
312 		    (tiled_x + copy_width == (1 << 14) ||
313 		     tiled_y + copy_height == (1 << 14)))
314 			return false;
315 
316 		/* The hw can read outside of the given linear buffer bounds,
317 		 * or access those pages but not touch the memory in case
318 		 * of writes. (it still causes a VM fault)
319 		 *
320 		 * Out-of-bounds memory access or page directory access must
321 		 * be prevented.
322 		 */
323 		int64_t start_linear_address, end_linear_address;
324 		unsigned granularity;
325 
326 		/* Deduce the size of reads from the linear surface. */
327 		switch (tiled_micro_mode) {
328 		case V_009910_ADDR_SURF_DISPLAY_MICRO_TILING:
329 			granularity = bpp == 1 ? 64 / (8*bpp) :
330 						 128 / (8*bpp);
331 			break;
332 		case V_009910_ADDR_SURF_THIN_MICRO_TILING:
333 		case V_009910_ADDR_SURF_DEPTH_MICRO_TILING:
334 			if (0 /* TODO: THICK microtiling */)
335 				granularity = bpp == 1 ? 32 / (8*bpp) :
336 					      bpp == 2 ? 64 / (8*bpp) :
337 					      bpp <= 8 ? 128 / (8*bpp) :
338 							 256 / (8*bpp);
339 			else
340 				granularity = bpp <= 2 ? 64 / (8*bpp) :
341 					      bpp <= 8 ? 128 / (8*bpp) :
342 							 256 / (8*bpp);
343 			break;
344 		default:
345 			return false;
346 		}
347 
348 		/* The linear reads start at tiled_x & ~(granularity - 1).
349 		 * If linear_x == 0 && tiled_x % granularity != 0, the hw
350 		 * starts reading from an address preceding linear_address!!!
351 		 */
352 		start_linear_address =
353 			linear->surface.level[linear_level].offset +
354 			bpp * (linear_z * linear_slice_pitch +
355 			       linear_y * linear_pitch +
356 			       linear_x);
357 		start_linear_address -= (int)(bpp * (tiled_x % granularity));
358 
359 		end_linear_address =
360 			linear->surface.level[linear_level].offset +
361 			bpp * ((linear_z + copy_depth - 1) * linear_slice_pitch +
362 			       (linear_y + copy_height - 1) * linear_pitch +
363 			       (linear_x + copy_width));
364 
365 		if ((tiled_x + copy_width) % granularity)
366 			end_linear_address += granularity -
367 					      (tiled_x + copy_width) % granularity;
368 
369 		if (start_linear_address < 0 ||
370 		    end_linear_address > linear->surface.surf_size)
371 			return false;
372 
373 		/* Check requirements. */
374 		if (tiled_address % 256 == 0 &&
375 		    linear_address % 4 == 0 &&
376 		    linear_pitch % xalign == 0 &&
377 		    linear_x % xalign == 0 &&
378 		    tiled_x % xalign == 0 &&
379 		    copy_width_aligned % xalign == 0 &&
380 		    tiled_micro_mode != V_009910_ADDR_SURF_ROTATED_MICRO_TILING &&
381 		    /* check if everything fits into the bitfields */
382 		    tiled->surface.tile_split <= 4096 &&
383 		    pitch_tile_max < (1 << 11) &&
384 		    slice_tile_max < (1 << 22) &&
385 		    linear_pitch <= (1 << 14) &&
386 		    linear_slice_pitch <= (1 << 28) &&
387 		    copy_width_aligned <= (1 << 14) &&
388 		    copy_height <= (1 << 14) &&
389 		    copy_depth <= (1 << 11)) {
390 			struct radeon_winsys_cs *cs = sctx->b.dma.cs;
391 			uint32_t direction = linear == rdst ? 1u << 31 : 0;
392 
393 			r600_need_dma_space(&sctx->b, 14, &rdst->resource, &rsrc->resource);
394 
395 			radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
396 							CIK_SDMA_COPY_SUB_OPCODE_TILED_SUB_WINDOW, 0) |
397 					direction);
398 			radeon_emit(cs, tiled_address);
399 			radeon_emit(cs, tiled_address >> 32);
400 			radeon_emit(cs, tiled_x | (tiled_y << 16));
401 			radeon_emit(cs, tiled_z | (pitch_tile_max << 16));
402 			radeon_emit(cs, slice_tile_max);
403 			radeon_emit(cs, encode_tile_info(sctx, tiled, tiled_level, true));
404 			radeon_emit(cs, linear_address);
405 			radeon_emit(cs, linear_address >> 32);
406 			radeon_emit(cs, linear_x | (linear_y << 16));
407 			radeon_emit(cs, linear_z | ((linear_pitch - 1) << 16));
408 			radeon_emit(cs, linear_slice_pitch - 1);
409 			if (sctx->b.chip_class == CIK) {
410 				radeon_emit(cs, copy_width_aligned | (copy_height << 16));
411 				radeon_emit(cs, copy_depth);
412 			} else {
413 				radeon_emit(cs, (copy_width_aligned - 1) | ((copy_height - 1) << 16));
414 				radeon_emit(cs, (copy_depth - 1));
415 			}
416 			return true;
417 		}
418 	}
419 
420 	/* Tiled -> Tiled sub-window copy. */
421 	if (dst_mode >= RADEON_SURF_MODE_1D &&
422 	    src_mode >= RADEON_SURF_MODE_1D &&
423 	    /* check if these fit into the bitfields */
424 	    src_address % 256 == 0 &&
425 	    dst_address % 256 == 0 &&
426 	    rsrc->surface.tile_split <= 4096 &&
427 	    rdst->surface.tile_split <= 4096 &&
428 	    dstx % 8 == 0 &&
429 	    dsty % 8 == 0 &&
430 	    srcx % 8 == 0 &&
431 	    srcy % 8 == 0 &&
432 	    /* this can either be equal, or display->rotated (VI only) */
433 	    (src_micro_mode == dst_micro_mode ||
434 	     (sctx->b.chip_class == VI &&
435 	      src_micro_mode == V_009910_ADDR_SURF_DISPLAY_MICRO_TILING &&
436 	      dst_micro_mode == V_009910_ADDR_SURF_ROTATED_MICRO_TILING))) {
437 		assert(src_pitch % 8 == 0);
438 		assert(dst_pitch % 8 == 0);
439 		assert(src_slice_pitch % 64 == 0);
440 		assert(dst_slice_pitch % 64 == 0);
441 		unsigned src_pitch_tile_max = src_pitch / 8 - 1;
442 		unsigned dst_pitch_tile_max = dst_pitch / 8 - 1;
443 		unsigned src_slice_tile_max = src_slice_pitch / 64 - 1;
444 		unsigned dst_slice_tile_max = dst_slice_pitch / 64 - 1;
445 		unsigned copy_width_aligned = copy_width;
446 		unsigned copy_height_aligned = copy_height;
447 
448 		/* If the region ends at the last pixel and is unaligned, we
449 		 * can copy the remainder of the tile that is not visible to
450 		 * make it aligned.
451 		 */
452 		if (copy_width % 8 != 0 &&
453 		    srcx + copy_width == src_width &&
454 		    dstx + copy_width == dst_width)
455 			copy_width_aligned = align(copy_width, 8);
456 
457 		if (copy_height % 8 != 0 &&
458 		    srcy + copy_height == src_height &&
459 		    dsty + copy_height == dst_height)
460 			copy_height_aligned = align(copy_height, 8);
461 
462 		/* check if these fit into the bitfields */
463 		if (src_pitch_tile_max < (1 << 11) &&
464 		    dst_pitch_tile_max < (1 << 11) &&
465 		    src_slice_tile_max < (1 << 22) &&
466 		    dst_slice_tile_max < (1 << 22) &&
467 		    copy_width_aligned <= (1 << 14) &&
468 		    copy_height_aligned <= (1 << 14) &&
469 		    copy_depth <= (1 << 11) &&
470 		    copy_width_aligned % 8 == 0 &&
471 		    copy_height_aligned % 8 == 0 &&
472 		    /* HW limitation - CIK: */
473 		    (sctx->b.chip_class != CIK ||
474 		     (copy_width_aligned < (1 << 14) &&
475 		      copy_height_aligned < (1 << 14) &&
476 		      copy_depth < (1 << 11))) &&
477 		    /* HW limitation - some CIK parts: */
478 		    ((sctx->b.family != CHIP_BONAIRE &&
479 		      sctx->b.family != CHIP_KAVERI &&
480 		      sctx->b.family != CHIP_KABINI &&
481 		      sctx->b.family != CHIP_MULLINS) ||
482 		     (srcx + copy_width_aligned != (1 << 14) &&
483 		      srcy + copy_height_aligned != (1 << 14) &&
484 		      dstx + copy_width != (1 << 14)))) {
485 			struct radeon_winsys_cs *cs = sctx->b.dma.cs;
486 
487 			r600_need_dma_space(&sctx->b, 15, &rdst->resource, &rsrc->resource);
488 
489 			radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
490 							CIK_SDMA_COPY_SUB_OPCODE_T2T_SUB_WINDOW, 0));
491 			radeon_emit(cs, src_address);
492 			radeon_emit(cs, src_address >> 32);
493 			radeon_emit(cs, srcx | (srcy << 16));
494 			radeon_emit(cs, srcz | (src_pitch_tile_max << 16));
495 			radeon_emit(cs, src_slice_tile_max);
496 			radeon_emit(cs, encode_tile_info(sctx, rsrc, src_level, true));
497 			radeon_emit(cs, dst_address);
498 			radeon_emit(cs, dst_address >> 32);
499 			radeon_emit(cs, dstx | (dsty << 16));
500 			radeon_emit(cs, dstz | (dst_pitch_tile_max << 16));
501 			radeon_emit(cs, dst_slice_tile_max);
502 			radeon_emit(cs, encode_tile_info(sctx, rdst, dst_level, false));
503 			if (sctx->b.chip_class == CIK) {
504 				radeon_emit(cs, copy_width_aligned |
505 						(copy_height_aligned << 16));
506 				radeon_emit(cs, copy_depth);
507 			} else {
508 				radeon_emit(cs, (copy_width_aligned - 8) |
509 						((copy_height_aligned - 8) << 16));
510 				radeon_emit(cs, (copy_depth - 1));
511 			}
512 			return true;
513 		}
514 	}
515 
516 	return false;
517 }
518 
cik_sdma_copy(struct pipe_context * ctx,struct pipe_resource * dst,unsigned dst_level,unsigned dstx,unsigned dsty,unsigned dstz,struct pipe_resource * src,unsigned src_level,const struct pipe_box * src_box)519 static void cik_sdma_copy(struct pipe_context *ctx,
520 			  struct pipe_resource *dst,
521 			  unsigned dst_level,
522 			  unsigned dstx, unsigned dsty, unsigned dstz,
523 			  struct pipe_resource *src,
524 			  unsigned src_level,
525 			  const struct pipe_box *src_box)
526 {
527 	struct si_context *sctx = (struct si_context *)ctx;
528 
529 	if (!sctx->b.dma.cs)
530 		goto fallback;
531 
532 	if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
533 		cik_sdma_copy_buffer(sctx, dst, src, dstx, src_box->x, src_box->width);
534 		return;
535 	}
536 
537 	if (cik_sdma_copy_texture(sctx, dst, dst_level, dstx, dsty, dstz,
538 				  src, src_level, src_box))
539 		return;
540 
541 fallback:
542 	si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
543 				src, src_level, src_box);
544 }
545 
cik_init_sdma_functions(struct si_context * sctx)546 void cik_init_sdma_functions(struct si_context *sctx)
547 {
548 	sctx->b.dma_copy = cik_sdma_copy;
549 	sctx->b.dma_clear_buffer = cik_sdma_clear_buffer;
550 }
551