• 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 
25 #include "sid.h"
26 #include "si_pipe.h"
27 
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)28 static void cik_sdma_copy_buffer(struct si_context *ctx,
29 				 struct pipe_resource *dst,
30 				 struct pipe_resource *src,
31 				 uint64_t dst_offset,
32 				 uint64_t src_offset,
33 				 uint64_t size)
34 {
35 	struct radeon_winsys_cs *cs = ctx->b.dma.cs;
36 	unsigned i, ncopy, csize;
37 	struct r600_resource *rdst = r600_resource(dst);
38 	struct r600_resource *rsrc = r600_resource(src);
39 
40 	/* Mark the buffer range of destination as valid (initialized),
41 	 * so that transfer_map knows it should wait for the GPU when mapping
42 	 * that range. */
43 	util_range_add(&rdst->valid_buffer_range, dst_offset,
44 		       dst_offset + size);
45 
46 	dst_offset += rdst->gpu_address;
47 	src_offset += rsrc->gpu_address;
48 
49 	ncopy = DIV_ROUND_UP(size, CIK_SDMA_COPY_MAX_SIZE);
50 	si_need_dma_space(&ctx->b, ncopy * 7, rdst, rsrc);
51 
52 	for (i = 0; i < ncopy; i++) {
53 		csize = MIN2(size, CIK_SDMA_COPY_MAX_SIZE);
54 		radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
55 						CIK_SDMA_COPY_SUB_OPCODE_LINEAR,
56 						0));
57 		radeon_emit(cs, ctx->b.chip_class >= GFX9 ? csize - 1 : csize);
58 		radeon_emit(cs, 0); /* src/dst endian swap */
59 		radeon_emit(cs, src_offset);
60 		radeon_emit(cs, src_offset >> 32);
61 		radeon_emit(cs, dst_offset);
62 		radeon_emit(cs, dst_offset >> 32);
63 		dst_offset += csize;
64 		src_offset += csize;
65 		size -= csize;
66 	}
67 }
68 
cik_sdma_clear_buffer(struct pipe_context * ctx,struct pipe_resource * dst,uint64_t offset,uint64_t size,unsigned clear_value)69 static void cik_sdma_clear_buffer(struct pipe_context *ctx,
70 				  struct pipe_resource *dst,
71 				  uint64_t offset,
72 				  uint64_t size,
73 				  unsigned clear_value)
74 {
75 	struct si_context *sctx = (struct si_context *)ctx;
76 	struct radeon_winsys_cs *cs = sctx->b.dma.cs;
77 	unsigned i, ncopy, csize;
78 	struct r600_resource *rdst = r600_resource(dst);
79 
80 	if (!cs || offset % 4 != 0 || size % 4 != 0 ||
81 	    dst->flags & PIPE_RESOURCE_FLAG_SPARSE) {
82 		ctx->clear_buffer(ctx, dst, offset, size, &clear_value, 4);
83 		return;
84 	}
85 
86 	/* Mark the buffer range of destination as valid (initialized),
87 	 * so that transfer_map knows it should wait for the GPU when mapping
88 	 * that range. */
89 	util_range_add(&rdst->valid_buffer_range, offset, offset + size);
90 
91 	offset += rdst->gpu_address;
92 
93 	/* the same maximum size as for copying */
94 	ncopy = DIV_ROUND_UP(size, CIK_SDMA_COPY_MAX_SIZE);
95 	si_need_dma_space(&sctx->b, ncopy * 5, rdst, NULL);
96 
97 	for (i = 0; i < ncopy; i++) {
98 		csize = MIN2(size, CIK_SDMA_COPY_MAX_SIZE);
99 		radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_PACKET_CONSTANT_FILL, 0,
100 						0x8000 /* dword copy */));
101 		radeon_emit(cs, offset);
102 		radeon_emit(cs, offset >> 32);
103 		radeon_emit(cs, clear_value);
104 		radeon_emit(cs, sctx->b.chip_class >= GFX9 ? csize - 1 : csize);
105 		offset += csize;
106 		size -= csize;
107 	}
108 }
109 
minify_as_blocks(unsigned width,unsigned level,unsigned blk_w)110 static unsigned minify_as_blocks(unsigned width, unsigned level, unsigned blk_w)
111 {
112 	width = u_minify(width, level);
113 	return DIV_ROUND_UP(width, blk_w);
114 }
115 
encode_tile_info(struct si_context * sctx,struct r600_texture * tex,unsigned level,bool set_bpp)116 static unsigned encode_tile_info(struct si_context *sctx,
117 				 struct r600_texture *tex, unsigned level,
118 				 bool set_bpp)
119 {
120 	struct radeon_info *info = &sctx->screen->info;
121 	unsigned tile_index = tex->surface.u.legacy.tiling_index[level];
122 	unsigned macro_tile_index = tex->surface.u.legacy.macro_tile_index;
123 	unsigned tile_mode = info->si_tile_mode_array[tile_index];
124 	unsigned macro_tile_mode = info->cik_macrotile_mode_array[macro_tile_index];
125 
126 	return (set_bpp ? util_logbase2(tex->surface.bpe) : 0) |
127 		(G_009910_ARRAY_MODE(tile_mode) << 3) |
128 		(G_009910_MICRO_TILE_MODE_NEW(tile_mode) << 8) |
129 		/* Non-depth modes don't have TILE_SPLIT set. */
130 		((util_logbase2(tex->surface.u.legacy.tile_split >> 6)) << 11) |
131 		(G_009990_BANK_WIDTH(macro_tile_mode) << 15) |
132 		(G_009990_BANK_HEIGHT(macro_tile_mode) << 18) |
133 		(G_009990_NUM_BANKS(macro_tile_mode) << 21) |
134 		(G_009990_MACRO_TILE_ASPECT(macro_tile_mode) << 24) |
135 		(G_009910_PIPE_CONFIG(tile_mode) << 26);
136 }
137 
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)138 static bool cik_sdma_copy_texture(struct si_context *sctx,
139 				  struct pipe_resource *dst,
140 				  unsigned dst_level,
141 				  unsigned dstx, unsigned dsty, unsigned dstz,
142 				  struct pipe_resource *src,
143 				  unsigned src_level,
144 				  const struct pipe_box *src_box)
145 {
146 	struct radeon_info *info = &sctx->screen->info;
147 	struct r600_texture *rsrc = (struct r600_texture*)src;
148 	struct r600_texture *rdst = (struct r600_texture*)dst;
149 	unsigned bpp = rdst->surface.bpe;
150 	uint64_t dst_address = rdst->resource.gpu_address +
151 			       rdst->surface.u.legacy.level[dst_level].offset;
152 	uint64_t src_address = rsrc->resource.gpu_address +
153 			       rsrc->surface.u.legacy.level[src_level].offset;
154 	unsigned dst_mode = rdst->surface.u.legacy.level[dst_level].mode;
155 	unsigned src_mode = rsrc->surface.u.legacy.level[src_level].mode;
156 	unsigned dst_tile_index = rdst->surface.u.legacy.tiling_index[dst_level];
157 	unsigned src_tile_index = rsrc->surface.u.legacy.tiling_index[src_level];
158 	unsigned dst_tile_mode = info->si_tile_mode_array[dst_tile_index];
159 	unsigned src_tile_mode = info->si_tile_mode_array[src_tile_index];
160 	unsigned dst_micro_mode = G_009910_MICRO_TILE_MODE_NEW(dst_tile_mode);
161 	unsigned src_micro_mode = G_009910_MICRO_TILE_MODE_NEW(src_tile_mode);
162 	unsigned dst_tile_swizzle = dst_mode == RADEON_SURF_MODE_2D ?
163 					    rdst->surface.tile_swizzle : 0;
164 	unsigned src_tile_swizzle = src_mode == RADEON_SURF_MODE_2D ?
165 					    rsrc->surface.tile_swizzle : 0;
166 	unsigned dst_pitch = rdst->surface.u.legacy.level[dst_level].nblk_x;
167 	unsigned src_pitch = rsrc->surface.u.legacy.level[src_level].nblk_x;
168 	uint64_t dst_slice_pitch = ((uint64_t)rdst->surface.u.legacy.level[dst_level].slice_size_dw * 4) / bpp;
169 	uint64_t src_slice_pitch = ((uint64_t)rsrc->surface.u.legacy.level[src_level].slice_size_dw * 4) / bpp;
170 	unsigned dst_width = minify_as_blocks(rdst->resource.b.b.width0,
171 					      dst_level, rdst->surface.blk_w);
172 	unsigned src_width = minify_as_blocks(rsrc->resource.b.b.width0,
173 					      src_level, rsrc->surface.blk_w);
174 	unsigned dst_height = minify_as_blocks(rdst->resource.b.b.height0,
175 					       dst_level, rdst->surface.blk_h);
176 	unsigned src_height = minify_as_blocks(rsrc->resource.b.b.height0,
177 					       src_level, rsrc->surface.blk_h);
178 	unsigned srcx = src_box->x / rsrc->surface.blk_w;
179 	unsigned srcy = src_box->y / rsrc->surface.blk_h;
180 	unsigned srcz = src_box->z;
181 	unsigned copy_width = DIV_ROUND_UP(src_box->width, rsrc->surface.blk_w);
182 	unsigned copy_height = DIV_ROUND_UP(src_box->height, rsrc->surface.blk_h);
183 	unsigned copy_depth = src_box->depth;
184 
185 	assert(src_level <= src->last_level);
186 	assert(dst_level <= dst->last_level);
187 	assert(rdst->surface.u.legacy.level[dst_level].offset +
188 	       dst_slice_pitch * bpp * (dstz + src_box->depth) <=
189 	       rdst->resource.buf->size);
190 	assert(rsrc->surface.u.legacy.level[src_level].offset +
191 	       src_slice_pitch * bpp * (srcz + src_box->depth) <=
192 	       rsrc->resource.buf->size);
193 
194 	if (!si_prepare_for_dma_blit(&sctx->b, rdst, dst_level, dstx, dsty,
195 					dstz, rsrc, src_level, src_box))
196 		return false;
197 
198 	dstx /= rdst->surface.blk_w;
199 	dsty /= rdst->surface.blk_h;
200 
201 	if (srcx >= (1 << 14) ||
202 	    srcy >= (1 << 14) ||
203 	    srcz >= (1 << 11) ||
204 	    dstx >= (1 << 14) ||
205 	    dsty >= (1 << 14) ||
206 	    dstz >= (1 << 11))
207 		return false;
208 
209 	dst_address |= dst_tile_swizzle << 8;
210 	src_address |= src_tile_swizzle << 8;
211 
212 	/* Linear -> linear sub-window copy. */
213 	if (dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED &&
214 	    src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED &&
215 	    /* check if everything fits into the bitfields */
216 	    src_pitch <= (1 << 14) &&
217 	    dst_pitch <= (1 << 14) &&
218 	    src_slice_pitch <= (1 << 28) &&
219 	    dst_slice_pitch <= (1 << 28) &&
220 	    copy_width <= (1 << 14) &&
221 	    copy_height <= (1 << 14) &&
222 	    copy_depth <= (1 << 11) &&
223 	    /* HW limitation - CIK: */
224 	    (sctx->b.chip_class != CIK ||
225 	     (copy_width < (1 << 14) &&
226 	      copy_height < (1 << 14) &&
227 	      copy_depth < (1 << 11))) &&
228 	    /* HW limitation - some CIK parts: */
229 	    ((sctx->b.family != CHIP_BONAIRE &&
230 	      sctx->b.family != CHIP_KAVERI) ||
231 	     (srcx + copy_width != (1 << 14) &&
232 	      srcy + copy_height != (1 << 14)))) {
233 		struct radeon_winsys_cs *cs = sctx->b.dma.cs;
234 
235 		si_need_dma_space(&sctx->b, 13, &rdst->resource, &rsrc->resource);
236 
237 		radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
238 						CIK_SDMA_COPY_SUB_OPCODE_LINEAR_SUB_WINDOW, 0) |
239 			    (util_logbase2(bpp) << 29));
240 		radeon_emit(cs, src_address);
241 		radeon_emit(cs, src_address >> 32);
242 		radeon_emit(cs, srcx | (srcy << 16));
243 		radeon_emit(cs, srcz | ((src_pitch - 1) << 16));
244 		radeon_emit(cs, src_slice_pitch - 1);
245 		radeon_emit(cs, dst_address);
246 		radeon_emit(cs, dst_address >> 32);
247 		radeon_emit(cs, dstx | (dsty << 16));
248 		radeon_emit(cs, dstz | ((dst_pitch - 1) << 16));
249 		radeon_emit(cs, dst_slice_pitch - 1);
250 		if (sctx->b.chip_class == CIK) {
251 			radeon_emit(cs, copy_width | (copy_height << 16));
252 			radeon_emit(cs, copy_depth);
253 		} else {
254 			radeon_emit(cs, (copy_width - 1) | ((copy_height - 1) << 16));
255 			radeon_emit(cs, (copy_depth - 1));
256 		}
257 		return true;
258 	}
259 
260 	/* Tiled <-> linear sub-window copy. */
261 	if ((src_mode >= RADEON_SURF_MODE_1D) != (dst_mode >= RADEON_SURF_MODE_1D)) {
262 		struct r600_texture *tiled = src_mode >= RADEON_SURF_MODE_1D ? rsrc : rdst;
263 		struct r600_texture *linear = tiled == rsrc ? rdst : rsrc;
264 		unsigned tiled_level =	tiled	== rsrc ? src_level : dst_level;
265 		unsigned linear_level =	linear	== rsrc ? src_level : dst_level;
266 		unsigned tiled_x =	tiled	== rsrc ? srcx : dstx;
267 		unsigned linear_x =	linear  == rsrc ? srcx : dstx;
268 		unsigned tiled_y =	tiled	== rsrc ? srcy : dsty;
269 		unsigned linear_y =	linear  == rsrc ? srcy : dsty;
270 		unsigned tiled_z =	tiled	== rsrc ? srcz : dstz;
271 		unsigned linear_z =	linear  == rsrc ? srcz : dstz;
272 		unsigned tiled_width =	tiled	== rsrc ? src_width : dst_width;
273 		unsigned linear_width =	linear	== rsrc ? src_width : dst_width;
274 		unsigned tiled_pitch =	tiled	== rsrc ? src_pitch : dst_pitch;
275 		unsigned linear_pitch =	linear	== rsrc ? src_pitch : dst_pitch;
276 		unsigned tiled_slice_pitch  = tiled  == rsrc ? src_slice_pitch : dst_slice_pitch;
277 		unsigned linear_slice_pitch = linear == rsrc ? src_slice_pitch : dst_slice_pitch;
278 		uint64_t tiled_address =  tiled  == rsrc ? src_address : dst_address;
279 		uint64_t linear_address = linear == rsrc ? src_address : dst_address;
280 		unsigned tiled_micro_mode = tiled == rsrc ? src_micro_mode : dst_micro_mode;
281 
282 		assert(tiled_pitch % 8 == 0);
283 		assert(tiled_slice_pitch % 64 == 0);
284 		unsigned pitch_tile_max = tiled_pitch / 8 - 1;
285 		unsigned slice_tile_max = tiled_slice_pitch / 64 - 1;
286 		unsigned xalign = MAX2(1, 4 / bpp);
287 		unsigned copy_width_aligned = copy_width;
288 
289 		/* If the region ends at the last pixel and is unaligned, we
290 		 * can copy the remainder of the line that is not visible to
291 		 * make it aligned.
292 		 */
293 		if (copy_width % xalign != 0 &&
294 		    linear_x + copy_width == linear_width &&
295 		    tiled_x  + copy_width == tiled_width &&
296 		    linear_x + align(copy_width, xalign) <= linear_pitch &&
297 		    tiled_x  + align(copy_width, xalign) <= tiled_pitch)
298 			copy_width_aligned = align(copy_width, xalign);
299 
300 		/* HW limitations. */
301 		if ((sctx->b.family == CHIP_BONAIRE ||
302 		     sctx->b.family == CHIP_KAVERI) &&
303 		    linear_pitch - 1 == 0x3fff &&
304 		    bpp == 16)
305 			return false;
306 
307 		if (sctx->b.chip_class == CIK &&
308 		    (copy_width_aligned == (1 << 14) ||
309 		     copy_height == (1 << 14) ||
310 		     copy_depth == (1 << 11)))
311 			return false;
312 
313 		if ((sctx->b.family == CHIP_BONAIRE ||
314 		     sctx->b.family == CHIP_KAVERI ||
315 		     sctx->b.family == CHIP_KABINI ||
316 		     sctx->b.family == CHIP_MULLINS) &&
317 		    (tiled_x + copy_width == (1 << 14) ||
318 		     tiled_y + copy_height == (1 << 14)))
319 			return false;
320 
321 		/* The hw can read outside of the given linear buffer bounds,
322 		 * or access those pages but not touch the memory in case
323 		 * of writes. (it still causes a VM fault)
324 		 *
325 		 * Out-of-bounds memory access or page directory access must
326 		 * be prevented.
327 		 */
328 		int64_t start_linear_address, end_linear_address;
329 		unsigned granularity;
330 
331 		/* Deduce the size of reads from the linear surface. */
332 		switch (tiled_micro_mode) {
333 		case V_009910_ADDR_SURF_DISPLAY_MICRO_TILING:
334 			granularity = bpp == 1 ? 64 / (8*bpp) :
335 						 128 / (8*bpp);
336 			break;
337 		case V_009910_ADDR_SURF_THIN_MICRO_TILING:
338 		case V_009910_ADDR_SURF_DEPTH_MICRO_TILING:
339 			if (0 /* TODO: THICK microtiling */)
340 				granularity = bpp == 1 ? 32 / (8*bpp) :
341 					      bpp == 2 ? 64 / (8*bpp) :
342 					      bpp <= 8 ? 128 / (8*bpp) :
343 							 256 / (8*bpp);
344 			else
345 				granularity = bpp <= 2 ? 64 / (8*bpp) :
346 					      bpp <= 8 ? 128 / (8*bpp) :
347 							 256 / (8*bpp);
348 			break;
349 		default:
350 			return false;
351 		}
352 
353 		/* The linear reads start at tiled_x & ~(granularity - 1).
354 		 * If linear_x == 0 && tiled_x % granularity != 0, the hw
355 		 * starts reading from an address preceding linear_address!!!
356 		 */
357 		start_linear_address =
358 			linear->surface.u.legacy.level[linear_level].offset +
359 			bpp * (linear_z * linear_slice_pitch +
360 			       linear_y * linear_pitch +
361 			       linear_x);
362 		start_linear_address -= (int)(bpp * (tiled_x % granularity));
363 
364 		end_linear_address =
365 			linear->surface.u.legacy.level[linear_level].offset +
366 			bpp * ((linear_z + copy_depth - 1) * linear_slice_pitch +
367 			       (linear_y + copy_height - 1) * linear_pitch +
368 			       (linear_x + copy_width));
369 
370 		if ((tiled_x + copy_width) % granularity)
371 			end_linear_address += granularity -
372 					      (tiled_x + copy_width) % granularity;
373 
374 		if (start_linear_address < 0 ||
375 		    end_linear_address > linear->surface.surf_size)
376 			return false;
377 
378 		/* Check requirements. */
379 		if (tiled_address % 256 == 0 &&
380 		    linear_address % 4 == 0 &&
381 		    linear_pitch % xalign == 0 &&
382 		    linear_x % xalign == 0 &&
383 		    tiled_x % xalign == 0 &&
384 		    copy_width_aligned % xalign == 0 &&
385 		    tiled_micro_mode != V_009910_ADDR_SURF_ROTATED_MICRO_TILING &&
386 		    /* check if everything fits into the bitfields */
387 		    tiled->surface.u.legacy.tile_split <= 4096 &&
388 		    pitch_tile_max < (1 << 11) &&
389 		    slice_tile_max < (1 << 22) &&
390 		    linear_pitch <= (1 << 14) &&
391 		    linear_slice_pitch <= (1 << 28) &&
392 		    copy_width_aligned <= (1 << 14) &&
393 		    copy_height <= (1 << 14) &&
394 		    copy_depth <= (1 << 11)) {
395 			struct radeon_winsys_cs *cs = sctx->b.dma.cs;
396 			uint32_t direction = linear == rdst ? 1u << 31 : 0;
397 
398 			si_need_dma_space(&sctx->b, 14, &rdst->resource, &rsrc->resource);
399 
400 			radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
401 							CIK_SDMA_COPY_SUB_OPCODE_TILED_SUB_WINDOW, 0) |
402 					direction);
403 			radeon_emit(cs, tiled_address);
404 			radeon_emit(cs, tiled_address >> 32);
405 			radeon_emit(cs, tiled_x | (tiled_y << 16));
406 			radeon_emit(cs, tiled_z | (pitch_tile_max << 16));
407 			radeon_emit(cs, slice_tile_max);
408 			radeon_emit(cs, encode_tile_info(sctx, tiled, tiled_level, true));
409 			radeon_emit(cs, linear_address);
410 			radeon_emit(cs, linear_address >> 32);
411 			radeon_emit(cs, linear_x | (linear_y << 16));
412 			radeon_emit(cs, linear_z | ((linear_pitch - 1) << 16));
413 			radeon_emit(cs, linear_slice_pitch - 1);
414 			if (sctx->b.chip_class == CIK) {
415 				radeon_emit(cs, copy_width_aligned | (copy_height << 16));
416 				radeon_emit(cs, copy_depth);
417 			} else {
418 				radeon_emit(cs, (copy_width_aligned - 1) | ((copy_height - 1) << 16));
419 				radeon_emit(cs, (copy_depth - 1));
420 			}
421 			return true;
422 		}
423 	}
424 
425 	/* Tiled -> Tiled sub-window copy. */
426 	if (dst_mode >= RADEON_SURF_MODE_1D &&
427 	    src_mode >= RADEON_SURF_MODE_1D &&
428 	    /* check if these fit into the bitfields */
429 	    src_address % 256 == 0 &&
430 	    dst_address % 256 == 0 &&
431 	    rsrc->surface.u.legacy.tile_split <= 4096 &&
432 	    rdst->surface.u.legacy.tile_split <= 4096 &&
433 	    dstx % 8 == 0 &&
434 	    dsty % 8 == 0 &&
435 	    srcx % 8 == 0 &&
436 	    srcy % 8 == 0 &&
437 	    /* this can either be equal, or display->rotated (VI+ only) */
438 	    (src_micro_mode == dst_micro_mode ||
439 	     (sctx->b.chip_class >= VI &&
440 	      src_micro_mode == V_009910_ADDR_SURF_DISPLAY_MICRO_TILING &&
441 	      dst_micro_mode == V_009910_ADDR_SURF_ROTATED_MICRO_TILING))) {
442 		assert(src_pitch % 8 == 0);
443 		assert(dst_pitch % 8 == 0);
444 		assert(src_slice_pitch % 64 == 0);
445 		assert(dst_slice_pitch % 64 == 0);
446 		unsigned src_pitch_tile_max = src_pitch / 8 - 1;
447 		unsigned dst_pitch_tile_max = dst_pitch / 8 - 1;
448 		unsigned src_slice_tile_max = src_slice_pitch / 64 - 1;
449 		unsigned dst_slice_tile_max = dst_slice_pitch / 64 - 1;
450 		unsigned copy_width_aligned = copy_width;
451 		unsigned copy_height_aligned = copy_height;
452 
453 		/* If the region ends at the last pixel and is unaligned, we
454 		 * can copy the remainder of the tile that is not visible to
455 		 * make it aligned.
456 		 */
457 		if (copy_width % 8 != 0 &&
458 		    srcx + copy_width == src_width &&
459 		    dstx + copy_width == dst_width)
460 			copy_width_aligned = align(copy_width, 8);
461 
462 		if (copy_height % 8 != 0 &&
463 		    srcy + copy_height == src_height &&
464 		    dsty + copy_height == dst_height)
465 			copy_height_aligned = align(copy_height, 8);
466 
467 		/* check if these fit into the bitfields */
468 		if (src_pitch_tile_max < (1 << 11) &&
469 		    dst_pitch_tile_max < (1 << 11) &&
470 		    src_slice_tile_max < (1 << 22) &&
471 		    dst_slice_tile_max < (1 << 22) &&
472 		    copy_width_aligned <= (1 << 14) &&
473 		    copy_height_aligned <= (1 << 14) &&
474 		    copy_depth <= (1 << 11) &&
475 		    copy_width_aligned % 8 == 0 &&
476 		    copy_height_aligned % 8 == 0 &&
477 		    /* HW limitation - CIK: */
478 		    (sctx->b.chip_class != CIK ||
479 		     (copy_width_aligned < (1 << 14) &&
480 		      copy_height_aligned < (1 << 14) &&
481 		      copy_depth < (1 << 11))) &&
482 		    /* HW limitation - some CIK parts: */
483 		    ((sctx->b.family != CHIP_BONAIRE &&
484 		      sctx->b.family != CHIP_KAVERI &&
485 		      sctx->b.family != CHIP_KABINI &&
486 		      sctx->b.family != CHIP_MULLINS) ||
487 		     (srcx + copy_width_aligned != (1 << 14) &&
488 		      srcy + copy_height_aligned != (1 << 14) &&
489 		      dstx + copy_width != (1 << 14)))) {
490 			struct radeon_winsys_cs *cs = sctx->b.dma.cs;
491 
492 			si_need_dma_space(&sctx->b, 15, &rdst->resource, &rsrc->resource);
493 
494 			radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
495 							CIK_SDMA_COPY_SUB_OPCODE_T2T_SUB_WINDOW, 0));
496 			radeon_emit(cs, src_address);
497 			radeon_emit(cs, src_address >> 32);
498 			radeon_emit(cs, srcx | (srcy << 16));
499 			radeon_emit(cs, srcz | (src_pitch_tile_max << 16));
500 			radeon_emit(cs, src_slice_tile_max);
501 			radeon_emit(cs, encode_tile_info(sctx, rsrc, src_level, true));
502 			radeon_emit(cs, dst_address);
503 			radeon_emit(cs, dst_address >> 32);
504 			radeon_emit(cs, dstx | (dsty << 16));
505 			radeon_emit(cs, dstz | (dst_pitch_tile_max << 16));
506 			radeon_emit(cs, dst_slice_tile_max);
507 			radeon_emit(cs, encode_tile_info(sctx, rdst, dst_level, false));
508 			if (sctx->b.chip_class == CIK) {
509 				radeon_emit(cs, copy_width_aligned |
510 						(copy_height_aligned << 16));
511 				radeon_emit(cs, copy_depth);
512 			} else {
513 				radeon_emit(cs, (copy_width_aligned - 8) |
514 						((copy_height_aligned - 8) << 16));
515 				radeon_emit(cs, (copy_depth - 1));
516 			}
517 			return true;
518 		}
519 	}
520 
521 	return false;
522 }
523 
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)524 static void cik_sdma_copy(struct pipe_context *ctx,
525 			  struct pipe_resource *dst,
526 			  unsigned dst_level,
527 			  unsigned dstx, unsigned dsty, unsigned dstz,
528 			  struct pipe_resource *src,
529 			  unsigned src_level,
530 			  const struct pipe_box *src_box)
531 {
532 	struct si_context *sctx = (struct si_context *)ctx;
533 
534 	if (!sctx->b.dma.cs ||
535 	    src->flags & PIPE_RESOURCE_FLAG_SPARSE ||
536 	    dst->flags & PIPE_RESOURCE_FLAG_SPARSE)
537 		goto fallback;
538 
539 	if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
540 		cik_sdma_copy_buffer(sctx, dst, src, dstx, src_box->x, src_box->width);
541 		return;
542 	}
543 
544 	if ((sctx->b.chip_class == CIK || sctx->b.chip_class == VI) &&
545 	    cik_sdma_copy_texture(sctx, dst, dst_level, dstx, dsty, dstz,
546 				  src, src_level, src_box))
547 		return;
548 
549 fallback:
550 	si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
551 				src, src_level, src_box);
552 }
553 
cik_init_sdma_functions(struct si_context * sctx)554 void cik_init_sdma_functions(struct si_context *sctx)
555 {
556 	sctx->b.dma_copy = cik_sdma_copy;
557 	sctx->b.dma_clear_buffer = cik_sdma_clear_buffer;
558 }
559