src/gallium/drivers/radeonsi/si_dma.c

   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
  32 static 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
  81 static void si_dma_copy_tile(struct si_context *ctx,
  82                              struct pipe_resource *dst,
  83                              unsigned dst_level,
  84                              unsigned dst_x,
  85                              unsigned dst_y,
  86                              unsigned dst_z,
  87                              struct pipe_resource *src,
  88                              unsigned src_level,
  89                              unsigned src_x,
  90                              unsigned src_y,
  91                              unsigned src_z,
  92                              unsigned copy_height,
  93                              unsigned pitch,
  94                              unsigned bpp)
  95 {
  96         struct radeon_winsys_cs *cs = ctx->b.dma.cs;
  97         struct r600_texture *rsrc = (struct r600_texture*)src;
  98         struct r600_texture *rdst = (struct r600_texture*)dst;
  99         unsigned dst_mode = rdst->surface.level[dst_level].mode;
 100         bool detile = dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED;
 101         struct r600_texture *rlinear = detile ? rdst : rsrc;
 102         struct r600_texture *rtiled = detile ? rsrc : rdst;
 103         unsigned linear_lvl = detile ? dst_level : src_level;
 104         unsigned tiled_lvl = detile ? src_level : dst_level;
 105         struct radeon_info *info = &ctx->screen->b.info;
 106         unsigned index = rtiled->surface.tiling_index[tiled_lvl];
 107         unsigned tile_mode = info->si_tile_mode_array[index];
 108         unsigned array_mode, lbpp, pitch_tile_max, slice_tile_max, size;
 109         unsigned ncopy, height, cheight, i;
 110         unsigned linear_x, linear_y, linear_z,  tiled_x, tiled_y, tiled_z;
 111         unsigned sub_cmd, bank_h, bank_w, mt_aspect, nbanks, tile_split, mt;
 112         uint64_t base, addr;
 113         unsigned pipe_config;
 114
 115         assert(dst_mode != rsrc->surface.level[src_level].mode);
 116
 117         sub_cmd = SI_DMA_COPY_TILED;
 118         lbpp = util_logbase2(bpp);
 119         pitch_tile_max = ((pitch / bpp) / 8) - 1;
 120
 121         linear_x = detile ? dst_x : src_x;
 122         linear_y = detile ? dst_y : src_y;
 123         linear_z = detile ? dst_z : src_z;
 124         tiled_x = detile ? src_x : dst_x;
 125         tiled_y = detile ? src_y : dst_y;
 126         tiled_z = detile ? src_z : dst_z;
 127
 128         assert(!util_format_is_depth_and_stencil(rtiled->resource.b.b.format));
 129
 130         array_mode = G_009910_ARRAY_MODE(tile_mode);
 131         slice_tile_max = (rtiled->surface.level[tiled_lvl].nblk_x *
 132                           rtiled->surface.level[tiled_lvl].nblk_y) / (8*8) - 1;
 133         /* linear height must be the same as the slice tile max height, it's ok even
 134          * if the linear destination/source have smaller heigh as the size of the
 135          * dma packet will be using the copy_height which is always smaller or equal
 136          * to the linear height
 137          */
 138         height = rtiled->surface.level[tiled_lvl].nblk_y;
 139         base = rtiled->surface.level[tiled_lvl].offset;
 140         addr = rlinear->surface.level[linear_lvl].offset;
 141         addr += rlinear->surface.level[linear_lvl].slice_size * linear_z;
 142         addr += linear_y * pitch + linear_x * bpp;
 143         bank_h = G_009910_BANK_HEIGHT(tile_mode);
 144         bank_w = G_009910_BANK_WIDTH(tile_mode);
 145         mt_aspect = G_009910_MACRO_TILE_ASPECT(tile_mode);
 146         /* Non-depth modes don't have TILE_SPLIT set. */
 147         tile_split = util_logbase2(rtiled->surface.tile_split >> 6);
 148         nbanks = G_009910_NUM_BANKS(tile_mode);
 149         base += rtiled->resource.gpu_address;
 150         addr += rlinear->resource.gpu_address;
 151
 152         pipe_config = G_009910_PIPE_CONFIG(tile_mode);
 153         mt = G_009910_MICRO_TILE_MODE(tile_mode);
 154         size = copy_height * pitch;
 155         ncopy = DIV_ROUND_UP(size, SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE);
 156         r600_need_dma_space(&ctx->b, ncopy * 9, &rdst->resource, &rsrc->resource);
 157
 158         for (i = 0; i < ncopy; i++) {
 159                 cheight = copy_height;
 160                 if (cheight * pitch > SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE) {
 161                         cheight = SI_DMA_COPY_MAX_DWORD_ALIGNED_SIZE / pitch;
 162                 }
 163                 size = cheight * pitch;
 164                 radeon_emit(cs, SI_DMA_PACKET(SI_DMA_PACKET_COPY, sub_cmd, size / 4));
 165                 radeon_emit(cs, base >> 8);
 166                 radeon_emit(cs, (detile << 31) | (array_mode << 27) |
 167                                 (lbpp << 24) | (bank_h << 21) |
 168                                 (bank_w << 18) | (mt_aspect << 16));
 169                 radeon_emit(cs, (pitch_tile_max << 0) | ((height - 1) << 16));
 170                 radeon_emit(cs, (slice_tile_max << 0) | (pipe_config << 26));
 171                 radeon_emit(cs, (tiled_x << 0) | (tiled_z << 18));
 172                 radeon_emit(cs, (tiled_y << 0) | (tile_split << 21) | (nbanks << 25) | (mt << 27));
 173                 radeon_emit(cs, addr & 0xfffffffc);
 174                 radeon_emit(cs, (addr >> 32UL) & 0xff);
 175                 copy_height -= cheight;
 176                 addr += cheight * pitch;
 177                 tiled_y += cheight;
 178         }
 179 }
 180
 181 static void si_dma_copy(struct pipe_context *ctx,
 182                         struct pipe_resource *dst,
 183                         unsigned dst_level,
 184                         unsigned dstx, unsigned dsty, unsigned dstz,
 185                         struct pipe_resource *src,
 186                         unsigned src_level,
 187                         const struct pipe_box *src_box)
 188 {
 189         struct si_context *sctx = (struct si_context *)ctx;
 190         struct r600_texture *rsrc = (struct r600_texture*)src;
 191         struct r600_texture *rdst = (struct r600_texture*)dst;
 192         unsigned dst_pitch, src_pitch, bpp, dst_mode, src_mode;
 193         unsigned src_w, dst_w;
 194         unsigned src_x, src_y;
 195         unsigned dst_x = dstx, dst_y = dsty, dst_z = dstz;
 196
 197         if (sctx->b.dma.cs == NULL) {
 198                 goto fallback;
 199         }
 200
 201         if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
 202                 si_dma_copy_buffer(sctx, dst, src, dst_x, src_box->x, src_box->width);
 203                 return;
 204         }
 205
 206         /* XXX: Using the asynchronous DMA engine for multi-dimensional
 207          * operations seems to cause random GPU lockups for various people.
 208          * While the root cause for this might need to be fixed in the kernel,
 209          * let's disable it for now.
 210          *
 211          * Before re-enabling this, please make sure you can hit all newly
 212          * enabled paths in your testing, preferably with both piglit and real
 213          * world apps, and get in touch with people on the bug reports below
 214          * for stability testing.
 215          *
 216          * https://bugs.freedesktop.org/show_bug.cgi?id=85647
 217          * https://bugs.freedesktop.org/show_bug.cgi?id=83500
 218          */
 219         goto fallback;
 220
 221         if (src_box->depth > 1 ||
 222             !r600_prepare_for_dma_blit(&sctx->b, rdst, dst_level, dstx, dsty,
 223                                         dstz, rsrc, src_level, src_box))
 224                 goto fallback;
 225
 226         src_x = util_format_get_nblocksx(src->format, src_box->x);
 227         dst_x = util_format_get_nblocksx(src->format, dst_x);
 228         src_y = util_format_get_nblocksy(src->format, src_box->y);
 229         dst_y = util_format_get_nblocksy(src->format, dst_y);
 230
 231         bpp = rdst->surface.bpe;
 232         dst_pitch = rdst->surface.level[dst_level].nblk_x * rdst->surface.bpe;
 233         src_pitch = rsrc->surface.level[src_level].nblk_x * rsrc->surface.bpe;
 234         src_w = u_minify(rsrc->resource.b.b.width0, src_level);
 235         dst_w = u_minify(rdst->resource.b.b.width0, dst_level);
 236
 237         dst_mode = rdst->surface.level[dst_level].mode;
 238         src_mode = rsrc->surface.level[src_level].mode;
 239
 240         if (src_pitch != dst_pitch || src_box->x || dst_x || src_w != dst_w ||
 241             src_box->width != src_w ||
 242             src_box->height != u_minify(rsrc->resource.b.b.height0, src_level) ||
 243             src_box->height != u_minify(rdst->resource.b.b.height0, dst_level) ||
 244             rsrc->surface.level[src_level].nblk_y !=
 245             rdst->surface.level[dst_level].nblk_y) {
 246                 /* FIXME si can do partial blit */
 247                 goto fallback;
 248         }
 249         /* the x test here are currently useless (because we don't support partial blit)
 250          * but keep them around so we don't forget about those
 251          */
 252         if ((src_pitch % 8) || (src_box->x % 8) || (dst_x % 8) ||
 253             (src_box->y % 8) || (dst_y % 8) || (src_box->height % 8)) {
 254                 goto fallback;
 255         }
 256
 257         if (src_mode == dst_mode) {
 258                 uint64_t dst_offset, src_offset;
 259                 /* simple dma blit would do NOTE code here assume :
 260                  *   src_box.x/y == 0
 261                  *   dst_x/y == 0
 262                  *   dst_pitch == src_pitch
 263                  */
 264                 src_offset= rsrc->surface.level[src_level].offset;
 265                 src_offset += rsrc->surface.level[src_level].slice_size * src_box->z;
 266                 src_offset += src_y * src_pitch + src_x * bpp;
 267                 dst_offset = rdst->surface.level[dst_level].offset;
 268                 dst_offset += rdst->surface.level[dst_level].slice_size * dst_z;
 269                 dst_offset += dst_y * dst_pitch + dst_x * bpp;
 270                 si_dma_copy_buffer(sctx, dst, src, dst_offset, src_offset,
 271                                    rsrc->surface.level[src_level].slice_size);
 272         } else {
 273                 si_dma_copy_tile(sctx, dst, dst_level, dst_x, dst_y, dst_z,
 274                                  src, src_level, src_x, src_y, src_box->z,
 275                                  src_box->height / rsrc->surface.blk_h,
 276                                  dst_pitch, bpp);
 277         }
 278         return;
 279
 280 fallback:
 281         si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
 282                                 src, src_level, src_box);
 283 }
 284
 285 void si_init_dma_functions(struct si_context *sctx)
 286 {
 287         sctx->b.dma_copy = si_dma_copy;
 288 }