1/*
2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
3 * Copyright 2018 Advanced Micro Devices, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * on the rights to use, copy, modify, merge, publish, distribute, sub
10 * license, and/or sell copies of the Software, and to permit persons to whom
11 * the Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
21 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
22 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
23 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26#include "sid.h"
27#include "si_pipe.h"
28
29#include "util/u_format.h"
30
31static void si_dma_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_cmdbuf *cs = ctx->dma_cs;
39	unsigned i, ncopy, count, max_size, sub_cmd, shift;
40	struct si_resource *sdst = si_resource(dst);
41	struct si_resource *ssrc = si_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(&sdst->valid_buffer_range, dst_offset,
47		       dst_offset + size);
48
49	dst_offset += sdst->gpu_address;
50	src_offset += ssrc->gpu_address;
51
52	/* see whether we should use the dword-aligned or byte-aligned copy */
53	if (!(dst_offset % 4) && !(src_offset % 4) && !(size % 4)) {
54		sub_cmd = SI_DMA_COPY_DWORD_ALIGNED;
55		shift = 2;
56		max_size = SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE;
57	} else {
58		sub_cmd = SI_DMA_COPY_BYTE_ALIGNED;
59		shift = 0;
60		max_size = SI_DMA_COPY_MAX_BYTE_ALIGNED_SIZE;
61	}
62
63	ncopy = DIV_ROUND_UP(size, max_size);
64	si_need_dma_space(ctx, ncopy * 5, sdst, ssrc);
65
66	for (i = 0; i < ncopy; i++) {
67		count = MIN2(size, max_size);
68		radeon_emit(cs, SI_DMA_PACKET(SI_DMA_PACKET_COPY, sub_cmd,
69					      count >> shift));
70		radeon_emit(cs, dst_offset);
71		radeon_emit(cs, src_offset);
72		radeon_emit(cs, (dst_offset >> 32UL) & 0xff);
73		radeon_emit(cs, (src_offset >> 32UL) & 0xff);
74		dst_offset += count;
75		src_offset += count;
76		size -= count;
77	}
78}
79
80static void si_dma_copy_tile(struct si_context *ctx,
81			     struct pipe_resource *dst,
82			     unsigned dst_level,
83			     unsigned dst_x,
84			     unsigned dst_y,
85			     unsigned dst_z,
86			     struct pipe_resource *src,
87			     unsigned src_level,
88			     unsigned src_x,
89			     unsigned src_y,
90			     unsigned src_z,
91			     unsigned copy_height,
92			     unsigned pitch,
93			     unsigned bpp)
94{
95	struct radeon_cmdbuf *cs = ctx->dma_cs;
96	struct si_texture *ssrc = (struct si_texture*)src;
97	struct si_texture *sdst = (struct si_texture*)dst;
98	unsigned dst_mode = sdst->surface.u.legacy.level[dst_level].mode;
99	bool detile = dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED;
100	struct si_texture *linear = detile ? sdst : ssrc;
101	struct si_texture *tiled = detile ? ssrc : sdst;
102	unsigned linear_lvl = detile ? dst_level : src_level;
103	unsigned tiled_lvl = detile ? src_level : dst_level;
104	struct radeon_info *info = &ctx->screen->info;
105	unsigned index = tiled->surface.u.legacy.tiling_index[tiled_lvl];
106	unsigned tile_mode = info->si_tile_mode_array[index];
107	unsigned array_mode, lbpp, pitch_tile_max, slice_tile_max, size;
108	unsigned ncopy, height, cheight, i;
109	unsigned linear_x, linear_y, linear_z,  tiled_x, tiled_y, tiled_z;
110	unsigned sub_cmd, bank_h, bank_w, mt_aspect, nbanks, tile_split, mt;
111	uint64_t base, addr;
112	unsigned pipe_config;
113
114	assert(dst_mode != ssrc->surface.u.legacy.level[src_level].mode);
115
116	sub_cmd = SI_DMA_COPY_TILED;
117	lbpp = util_logbase2(bpp);
118	pitch_tile_max = ((pitch / bpp) / 8) - 1;
119
120	linear_x = detile ? dst_x : src_x;
121	linear_y = detile ? dst_y : src_y;
122	linear_z = detile ? dst_z : src_z;
123	tiled_x = detile ? src_x : dst_x;
124	tiled_y = detile ? src_y : dst_y;
125	tiled_z = detile ? src_z : dst_z;
126
127	assert(!util_format_is_depth_and_stencil(tiled->buffer.b.b.format));
128
129	array_mode = G_009910_ARRAY_MODE(tile_mode);
130	slice_tile_max = (tiled->surface.u.legacy.level[tiled_lvl].nblk_x *
131			  tiled->surface.u.legacy.level[tiled_lvl].nblk_y) / (8*8) - 1;
132	/* linear height must be the same as the slice tile max height, it's ok even
133	 * if the linear destination/source have smaller heigh as the size of the
134	 * dma packet will be using the copy_height which is always smaller or equal
135	 * to the linear height
136	 */
137	height = tiled->surface.u.legacy.level[tiled_lvl].nblk_y;
138	base = tiled->surface.u.legacy.level[tiled_lvl].offset;
139	addr = linear->surface.u.legacy.level[linear_lvl].offset;
140	addr += (uint64_t)linear->surface.u.legacy.level[linear_lvl].slice_size_dw * 4 * linear_z;
141	addr += linear_y * pitch + linear_x * bpp;
142	bank_h = G_009910_BANK_HEIGHT(tile_mode);
143	bank_w = G_009910_BANK_WIDTH(tile_mode);
144	mt_aspect = G_009910_MACRO_TILE_ASPECT(tile_mode);
145	/* Non-depth modes don't have TILE_SPLIT set. */
146	tile_split = util_logbase2(tiled->surface.u.legacy.tile_split >> 6);
147	nbanks = G_009910_NUM_BANKS(tile_mode);
148	base += tiled->buffer.gpu_address;
149	addr += linear->buffer.gpu_address;
150
151	pipe_config = G_009910_PIPE_CONFIG(tile_mode);
152	mt = G_009910_MICRO_TILE_MODE(tile_mode);
153	size = copy_height * pitch;
154	ncopy = DIV_ROUND_UP(size, SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE);
155	si_need_dma_space(ctx, ncopy * 9, &sdst->buffer, &ssrc->buffer);
156
157	for (i = 0; i < ncopy; i++) {
158		cheight = copy_height;
159		if (cheight * pitch > SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE) {
160			cheight = SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE / pitch;
161		}
162		size = cheight * pitch;
163		radeon_emit(cs, SI_DMA_PACKET(SI_DMA_PACKET_COPY, sub_cmd, size / 4));
164		radeon_emit(cs, base >> 8);
165		radeon_emit(cs, (detile << 31) | (array_mode << 27) |
166				(lbpp << 24) | (bank_h << 21) |
167				(bank_w << 18) | (mt_aspect << 16));
168		radeon_emit(cs, (pitch_tile_max << 0) | ((height - 1) << 16));
169		radeon_emit(cs, (slice_tile_max << 0) | (pipe_config << 26));
170		radeon_emit(cs, (tiled_x << 0) | (tiled_z << 18));
171		radeon_emit(cs, (tiled_y << 0) | (tile_split << 21) | (nbanks << 25) | (mt << 27));
172		radeon_emit(cs, addr & 0xfffffffc);
173		radeon_emit(cs, (addr >> 32UL) & 0xff);
174		copy_height -= cheight;
175		addr += cheight * pitch;
176		tiled_y += cheight;
177	}
178}
179
180static void si_dma_copy(struct pipe_context *ctx,
181			struct pipe_resource *dst,
182			unsigned dst_level,
183			unsigned dstx, unsigned dsty, unsigned dstz,
184			struct pipe_resource *src,
185			unsigned src_level,
186			const struct pipe_box *src_box)
187{
188	struct si_context *sctx = (struct si_context *)ctx;
189	struct si_texture *ssrc = (struct si_texture*)src;
190	struct si_texture *sdst = (struct si_texture*)dst;
191	unsigned dst_pitch, src_pitch, bpp, dst_mode, src_mode;
192	unsigned src_w, dst_w;
193	unsigned src_x, src_y;
194	unsigned dst_x = dstx, dst_y = dsty, dst_z = dstz;
195
196	if (sctx->dma_cs == NULL ||
197	    src->flags & PIPE_RESOURCE_FLAG_SPARSE ||
198	    dst->flags & PIPE_RESOURCE_FLAG_SPARSE) {
199		goto fallback;
200	}
201
202	if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
203		si_dma_copy_buffer(sctx, dst, src, dst_x, src_box->x, src_box->width);
204		return;
205	}
206
207	/* XXX: Using the asynchronous DMA engine for multi-dimensional
208	 * operations seems to cause random GPU lockups for various people.
209	 * While the root cause for this might need to be fixed in the kernel,
210	 * let's disable it for now.
211	 *
212	 * Before re-enabling this, please make sure you can hit all newly
213	 * enabled paths in your testing, preferably with both piglit and real
214	 * world apps, and get in touch with people on the bug reports below
215	 * for stability testing.
216	 *
217	 * https://bugs.freedesktop.org/show_bug.cgi?id=85647
218	 * https://bugs.freedesktop.org/show_bug.cgi?id=83500
219	 */
220	goto fallback;
221
222	if (src_box->depth > 1 ||
223	    !si_prepare_for_dma_blit(sctx, sdst, dst_level, dstx, dsty,
224				     dstz, ssrc, src_level, src_box))
225		goto fallback;
226
227	src_x = util_format_get_nblocksx(src->format, src_box->x);
228	dst_x = util_format_get_nblocksx(src->format, dst_x);
229	src_y = util_format_get_nblocksy(src->format, src_box->y);
230	dst_y = util_format_get_nblocksy(src->format, dst_y);
231
232	bpp = sdst->surface.bpe;
233	dst_pitch = sdst->surface.u.legacy.level[dst_level].nblk_x * sdst->surface.bpe;
234	src_pitch = ssrc->surface.u.legacy.level[src_level].nblk_x * ssrc->surface.bpe;
235	src_w = u_minify(ssrc->buffer.b.b.width0, src_level);
236	dst_w = u_minify(sdst->buffer.b.b.width0, dst_level);
237
238	dst_mode = sdst->surface.u.legacy.level[dst_level].mode;
239	src_mode = ssrc->surface.u.legacy.level[src_level].mode;
240
241	if (src_pitch != dst_pitch || src_box->x || dst_x || src_w != dst_w ||
242	    src_box->width != src_w ||
243	    src_box->height != u_minify(ssrc->buffer.b.b.height0, src_level) ||
244	    src_box->height != u_minify(sdst->buffer.b.b.height0, dst_level) ||
245	    ssrc->surface.u.legacy.level[src_level].nblk_y !=
246	    sdst->surface.u.legacy.level[dst_level].nblk_y) {
247		/* FIXME si can do partial blit */
248		goto fallback;
249	}
250	/* the x test here are currently useless (because we don't support partial blit)
251	 * but keep them around so we don't forget about those
252	 */
253	if ((src_pitch % 8) || (src_box->x % 8) || (dst_x % 8) ||
254	    (src_box->y % 8) || (dst_y % 8) || (src_box->height % 8)) {
255		goto fallback;
256	}
257
258	if (src_mode == dst_mode) {
259		uint64_t dst_offset, src_offset;
260		/* simple dma blit would do NOTE code here assume :
261		 *   src_box.x/y == 0
262		 *   dst_x/y == 0
263		 *   dst_pitch == src_pitch
264		 */
265		src_offset= ssrc->surface.u.legacy.level[src_level].offset;
266		src_offset += (uint64_t)ssrc->surface.u.legacy.level[src_level].slice_size_dw * 4 * src_box->z;
267		src_offset += src_y * src_pitch + src_x * bpp;
268		dst_offset = sdst->surface.u.legacy.level[dst_level].offset;
269		dst_offset += (uint64_t)sdst->surface.u.legacy.level[dst_level].slice_size_dw * 4 * dst_z;
270		dst_offset += dst_y * dst_pitch + dst_x * bpp;
271		si_dma_copy_buffer(sctx, dst, src, dst_offset, src_offset,
272				   (uint64_t)ssrc->surface.u.legacy.level[src_level].slice_size_dw * 4);
273	} else {
274		si_dma_copy_tile(sctx, dst, dst_level, dst_x, dst_y, dst_z,
275				 src, src_level, src_x, src_y, src_box->z,
276				 src_box->height / ssrc->surface.blk_h,
277				 dst_pitch, bpp);
278	}
279	return;
280
281fallback:
282	si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
283				src, src_level, src_box);
284}
285
286void si_init_dma_functions(struct si_context *sctx)
287{
288	sctx->dma_copy = si_dma_copy;
289}
290