1/*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 *
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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 *      Jerome Glisse
25 */
26
27#include "sid.h"
28#include "si_pipe.h"
29
30#include "util/u_format.h"
31
32static void si_dma_copy_buffer(struct si_context *ctx,
33				struct pipe_resource *dst,
34				struct pipe_resource *src,
35				uint64_t dst_offset,
36				uint64_t src_offset,
37				uint64_t size)
38{
39	struct radeon_winsys_cs *cs = ctx->b.dma.cs;
40	unsigned i, ncopy, count, max_size, sub_cmd, shift;
41	struct r600_resource *rdst = (struct r600_resource*)dst;
42	struct r600_resource *rsrc = (struct r600_resource*)src;
43
44	/* Mark the buffer range of destination as valid (initialized),
45	 * so that transfer_map knows it should wait for the GPU when mapping
46	 * that range. */
47	util_range_add(&rdst->valid_buffer_range, dst_offset,
48		       dst_offset + size);
49
50	dst_offset += rdst->gpu_address;
51	src_offset += rsrc->gpu_address;
52
53	/* see whether we should use the dword-aligned or byte-aligned copy */
54	if (!(dst_offset % 4) && !(src_offset % 4) && !(size % 4)) {
55		sub_cmd = SI_DMA_COPY_DWORD_ALIGNED;
56		shift = 2;
57		max_size = SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE;
58	} else {
59		sub_cmd = SI_DMA_COPY_BYTE_ALIGNED;
60		shift = 0;
61		max_size = SI_DMA_COPY_MAX_BYTE_ALIGNED_SIZE;
62	}
63
64	ncopy = DIV_ROUND_UP(size, max_size);
65	r600_need_dma_space(&ctx->b, ncopy * 5, rdst, rsrc);
66
67	for (i = 0; i < ncopy; i++) {
68		count = MIN2(size, max_size);
69		radeon_emit(cs, SI_DMA_PACKET(SI_DMA_PACKET_COPY, sub_cmd,
70					      count >> shift));
71		radeon_emit(cs, dst_offset);
72		radeon_emit(cs, src_offset);
73		radeon_emit(cs, (dst_offset >> 32UL) & 0xff);
74		radeon_emit(cs, (src_offset >> 32UL) & 0xff);
75		dst_offset += count;
76		src_offset += count;
77		size -= count;
78	}
79}
80
81static void si_dma_clear_buffer(struct pipe_context *ctx,
82				struct pipe_resource *dst,
83				uint64_t offset,
84				uint64_t size,
85				unsigned clear_value)
86{
87	struct si_context *sctx = (struct si_context *)ctx;
88	struct radeon_winsys_cs *cs = sctx->b.dma.cs;
89	unsigned i, ncopy, csize;
90	struct r600_resource *rdst = r600_resource(dst);
91
92	if (!cs || offset % 4 != 0 || size % 4 != 0) {
93		ctx->clear_buffer(ctx, dst, offset, size, &clear_value, 4);
94		return;
95	}
96
97	/* Mark the buffer range of destination as valid (initialized),
98	 * so that transfer_map knows it should wait for the GPU when mapping
99	 * that range. */
100	util_range_add(&rdst->valid_buffer_range, offset, offset + size);
101
102	offset += rdst->gpu_address;
103
104	/* the same maximum size as for copying */
105	ncopy = DIV_ROUND_UP(size, SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE);
106	r600_need_dma_space(&sctx->b, ncopy * 4, rdst, NULL);
107
108	for (i = 0; i < ncopy; i++) {
109		csize = MIN2(size, SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE);
110		radeon_emit(cs, SI_DMA_PACKET(SI_DMA_PACKET_CONSTANT_FILL, 0,
111					      csize / 4));
112		radeon_emit(cs, offset);
113		radeon_emit(cs, clear_value);
114		radeon_emit(cs, (offset >> 32) << 16);
115		offset += csize;
116		size -= csize;
117	}
118}
119
120static void si_dma_copy_tile(struct si_context *ctx,
121			     struct pipe_resource *dst,
122			     unsigned dst_level,
123			     unsigned dst_x,
124			     unsigned dst_y,
125			     unsigned dst_z,
126			     struct pipe_resource *src,
127			     unsigned src_level,
128			     unsigned src_x,
129			     unsigned src_y,
130			     unsigned src_z,
131			     unsigned copy_height,
132			     unsigned pitch,
133			     unsigned bpp)
134{
135	struct radeon_winsys_cs *cs = ctx->b.dma.cs;
136	struct r600_texture *rsrc = (struct r600_texture*)src;
137	struct r600_texture *rdst = (struct r600_texture*)dst;
138	unsigned dst_mode = rdst->surface.level[dst_level].mode;
139	bool detile = dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED;
140	struct r600_texture *rlinear = detile ? rdst : rsrc;
141	struct r600_texture *rtiled = detile ? rsrc : rdst;
142	unsigned linear_lvl = detile ? dst_level : src_level;
143	unsigned tiled_lvl = detile ? src_level : dst_level;
144	struct radeon_info *info = &ctx->screen->b.info;
145	unsigned index = rtiled->surface.tiling_index[tiled_lvl];
146	unsigned tile_mode = info->si_tile_mode_array[index];
147	unsigned array_mode, lbpp, pitch_tile_max, slice_tile_max, size;
148	unsigned ncopy, height, cheight, i;
149	unsigned linear_x, linear_y, linear_z,  tiled_x, tiled_y, tiled_z;
150	unsigned sub_cmd, bank_h, bank_w, mt_aspect, nbanks, tile_split, mt;
151	uint64_t base, addr;
152	unsigned pipe_config;
153
154	assert(dst_mode != rsrc->surface.level[src_level].mode);
155
156	sub_cmd = SI_DMA_COPY_TILED;
157	lbpp = util_logbase2(bpp);
158	pitch_tile_max = ((pitch / bpp) / 8) - 1;
159
160	linear_x = detile ? dst_x : src_x;
161	linear_y = detile ? dst_y : src_y;
162	linear_z = detile ? dst_z : src_z;
163	tiled_x = detile ? src_x : dst_x;
164	tiled_y = detile ? src_y : dst_y;
165	tiled_z = detile ? src_z : dst_z;
166
167	assert(!util_format_is_depth_and_stencil(rtiled->resource.b.b.format));
168
169	array_mode = G_009910_ARRAY_MODE(tile_mode);
170	slice_tile_max = (rtiled->surface.level[tiled_lvl].nblk_x *
171			  rtiled->surface.level[tiled_lvl].nblk_y) / (8*8) - 1;
172	/* linear height must be the same as the slice tile max height, it's ok even
173	 * if the linear destination/source have smaller heigh as the size of the
174	 * dma packet will be using the copy_height which is always smaller or equal
175	 * to the linear height
176	 */
177	height = rtiled->surface.level[tiled_lvl].nblk_y;
178	base = rtiled->surface.level[tiled_lvl].offset;
179	addr = rlinear->surface.level[linear_lvl].offset;
180	addr += rlinear->surface.level[linear_lvl].slice_size * linear_z;
181	addr += linear_y * pitch + linear_x * bpp;
182	bank_h = G_009910_BANK_HEIGHT(tile_mode);
183	bank_w = G_009910_BANK_WIDTH(tile_mode);
184	mt_aspect = G_009910_MACRO_TILE_ASPECT(tile_mode);
185	/* Non-depth modes don't have TILE_SPLIT set. */
186	tile_split = util_logbase2(rtiled->surface.tile_split >> 6);
187	nbanks = G_009910_NUM_BANKS(tile_mode);
188	base += rtiled->resource.gpu_address;
189	addr += rlinear->resource.gpu_address;
190
191	pipe_config = G_009910_PIPE_CONFIG(tile_mode);
192	mt = G_009910_MICRO_TILE_MODE(tile_mode);
193	size = copy_height * pitch;
194	ncopy = DIV_ROUND_UP(size, SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE);
195	r600_need_dma_space(&ctx->b, ncopy * 9, &rdst->resource, &rsrc->resource);
196
197	for (i = 0; i < ncopy; i++) {
198		cheight = copy_height;
199		if (cheight * pitch > SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE) {
200			cheight = SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE / pitch;
201		}
202		size = cheight * pitch;
203		radeon_emit(cs, SI_DMA_PACKET(SI_DMA_PACKET_COPY, sub_cmd, size / 4));
204		radeon_emit(cs, base >> 8);
205		radeon_emit(cs, (detile << 31) | (array_mode << 27) |
206				(lbpp << 24) | (bank_h << 21) |
207				(bank_w << 18) | (mt_aspect << 16));
208		radeon_emit(cs, (pitch_tile_max << 0) | ((height - 1) << 16));
209		radeon_emit(cs, (slice_tile_max << 0) | (pipe_config << 26));
210		radeon_emit(cs, (tiled_x << 0) | (tiled_z << 18));
211		radeon_emit(cs, (tiled_y << 0) | (tile_split << 21) | (nbanks << 25) | (mt << 27));
212		radeon_emit(cs, addr & 0xfffffffc);
213		radeon_emit(cs, (addr >> 32UL) & 0xff);
214		copy_height -= cheight;
215		addr += cheight * pitch;
216		tiled_y += cheight;
217	}
218}
219
220static void si_dma_copy(struct pipe_context *ctx,
221			struct pipe_resource *dst,
222			unsigned dst_level,
223			unsigned dstx, unsigned dsty, unsigned dstz,
224			struct pipe_resource *src,
225			unsigned src_level,
226			const struct pipe_box *src_box)
227{
228	struct si_context *sctx = (struct si_context *)ctx;
229	struct r600_texture *rsrc = (struct r600_texture*)src;
230	struct r600_texture *rdst = (struct r600_texture*)dst;
231	unsigned dst_pitch, src_pitch, bpp, dst_mode, src_mode;
232	unsigned src_w, dst_w;
233	unsigned src_x, src_y;
234	unsigned dst_x = dstx, dst_y = dsty, dst_z = dstz;
235
236	if (sctx->b.dma.cs == NULL) {
237		goto fallback;
238	}
239
240	if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
241		si_dma_copy_buffer(sctx, dst, src, dst_x, src_box->x, src_box->width);
242		return;
243	}
244
245	/* XXX: Using the asynchronous DMA engine for multi-dimensional
246	 * operations seems to cause random GPU lockups for various people.
247	 * While the root cause for this might need to be fixed in the kernel,
248	 * let's disable it for now.
249	 *
250	 * Before re-enabling this, please make sure you can hit all newly
251	 * enabled paths in your testing, preferably with both piglit and real
252	 * world apps, and get in touch with people on the bug reports below
253	 * for stability testing.
254	 *
255	 * https://bugs.freedesktop.org/show_bug.cgi?id=85647
256	 * https://bugs.freedesktop.org/show_bug.cgi?id=83500
257	 */
258	goto fallback;
259
260	if (src_box->depth > 1 ||
261	    !r600_prepare_for_dma_blit(&sctx->b, rdst, dst_level, dstx, dsty,
262					dstz, rsrc, src_level, src_box))
263		goto fallback;
264
265	src_x = util_format_get_nblocksx(src->format, src_box->x);
266	dst_x = util_format_get_nblocksx(src->format, dst_x);
267	src_y = util_format_get_nblocksy(src->format, src_box->y);
268	dst_y = util_format_get_nblocksy(src->format, dst_y);
269
270	bpp = rdst->surface.bpe;
271	dst_pitch = rdst->surface.level[dst_level].nblk_x * rdst->surface.bpe;
272	src_pitch = rsrc->surface.level[src_level].nblk_x * rsrc->surface.bpe;
273	src_w = u_minify(rsrc->resource.b.b.width0, src_level);
274	dst_w = u_minify(rdst->resource.b.b.width0, dst_level);
275
276	dst_mode = rdst->surface.level[dst_level].mode;
277	src_mode = rsrc->surface.level[src_level].mode;
278
279	if (src_pitch != dst_pitch || src_box->x || dst_x || src_w != dst_w ||
280	    src_box->width != src_w ||
281	    src_box->height != u_minify(rsrc->resource.b.b.height0, src_level) ||
282	    src_box->height != u_minify(rdst->resource.b.b.height0, dst_level) ||
283	    rsrc->surface.level[src_level].nblk_y !=
284	    rdst->surface.level[dst_level].nblk_y) {
285		/* FIXME si can do partial blit */
286		goto fallback;
287	}
288	/* the x test here are currently useless (because we don't support partial blit)
289	 * but keep them around so we don't forget about those
290	 */
291	if ((src_pitch % 8) || (src_box->x % 8) || (dst_x % 8) ||
292	    (src_box->y % 8) || (dst_y % 8) || (src_box->height % 8)) {
293		goto fallback;
294	}
295
296	if (src_mode == dst_mode) {
297		uint64_t dst_offset, src_offset;
298		/* simple dma blit would do NOTE code here assume :
299		 *   src_box.x/y == 0
300		 *   dst_x/y == 0
301		 *   dst_pitch == src_pitch
302		 */
303		src_offset= rsrc->surface.level[src_level].offset;
304		src_offset += rsrc->surface.level[src_level].slice_size * src_box->z;
305		src_offset += src_y * src_pitch + src_x * bpp;
306		dst_offset = rdst->surface.level[dst_level].offset;
307		dst_offset += rdst->surface.level[dst_level].slice_size * dst_z;
308		dst_offset += dst_y * dst_pitch + dst_x * bpp;
309		si_dma_copy_buffer(sctx, dst, src, dst_offset, src_offset,
310				   rsrc->surface.level[src_level].slice_size);
311	} else {
312		si_dma_copy_tile(sctx, dst, dst_level, dst_x, dst_y, dst_z,
313				 src, src_level, src_x, src_y, src_box->z,
314				 src_box->height / rsrc->surface.blk_h,
315				 dst_pitch, bpp);
316	}
317	return;
318
319fallback:
320	si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
321				src, src_level, src_box);
322}
323
324void si_init_dma_functions(struct si_context *sctx)
325{
326	sctx->b.dma_copy = si_dma_copy;
327	sctx->b.dma_clear_buffer = si_dma_clear_buffer;
328}
329