src/gallium/drivers/radeonsi/si_cp_dma.c

   1 /*
   2  * Copyright 2013 Advanced Micro Devices, Inc.
   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  *      Marek Olšák <maraeo@gmail.com>
  25  */
  26
  27 #include "si_pipe.h"
  28 #include "sid.h"
  29 #include "radeon/r600_cs.h"
  30
  31
  32 /* Set this if you want the 3D engine to wait until CP DMA is done.
  33  * It should be set on the last CP DMA packet. */
  34 #define R600_CP_DMA_SYNC        (1 << 0) /* R600+ */
  35
  36 /* Set this if the source data was used as a destination in a previous CP DMA
  37  * packet. It's for preventing a read-after-write (RAW) hazard between two
  38  * CP DMA packets. */
  39 #define SI_CP_DMA_RAW_WAIT      (1 << 1) /* SI+ */
  40 #define CIK_CP_DMA_USE_L2       (1 << 2)
  41
  42 /* Emit a CP DMA packet to do a copy from one buffer to another.
  43  * The size must fit in bits [20:0].
  44  */
  45 static void si_emit_cp_dma_copy_buffer(struct si_context *sctx,
  46                                        uint64_t dst_va, uint64_t src_va,
  47                                        unsigned size, unsigned flags)
  48 {
  49         struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
  50         uint32_t sync_flag = flags & R600_CP_DMA_SYNC ? S_411_CP_SYNC(1) : 0;
  51         uint32_t raw_wait = flags & SI_CP_DMA_RAW_WAIT ? S_414_RAW_WAIT(1) : 0;
  52         uint32_t sel = flags & CIK_CP_DMA_USE_L2 ?
  53                            S_411_SRC_SEL(V_411_SRC_ADDR_TC_L2) |
  54                            S_411_DSL_SEL(V_411_DST_ADDR_TC_L2) : 0;
  55
  56         assert(size);
  57         assert((size & ((1<<21)-1)) == size);
  58
  59         if (sctx->b.chip_class >= CIK) {
  60                 radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
  61                 radeon_emit(cs, sync_flag | sel);       /* CP_SYNC [31] */
  62                 radeon_emit(cs, src_va);                /* SRC_ADDR_LO [31:0] */
  63                 radeon_emit(cs, src_va >> 32);          /* SRC_ADDR_HI [31:0] */
  64                 radeon_emit(cs, dst_va);                /* DST_ADDR_LO [31:0] */
  65                 radeon_emit(cs, dst_va >> 32);          /* DST_ADDR_HI [31:0] */
  66                 radeon_emit(cs, size | raw_wait);       /* COMMAND [29:22] | BYTE_COUNT [20:0] */
  67         } else {
  68                 radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
  69                 radeon_emit(cs, src_va);                        /* SRC_ADDR_LO [31:0] */
  70                 radeon_emit(cs, sync_flag | ((src_va >> 32) & 0xffff)); /* CP_SYNC [31] | SRC_ADDR_HI [15:0] */
  71                 radeon_emit(cs, dst_va);                        /* DST_ADDR_LO [31:0] */
  72                 radeon_emit(cs, (dst_va >> 32) & 0xffff);       /* DST_ADDR_HI [15:0] */
  73                 radeon_emit(cs, size | raw_wait);               /* COMMAND [29:22] | BYTE_COUNT [20:0] */
  74         }
  75 }
  76
  77 /* Emit a CP DMA packet to clear a buffer. The size must fit in bits [20:0]. */
  78 static void si_emit_cp_dma_clear_buffer(struct si_context *sctx,
  79                                         uint64_t dst_va, unsigned size,
  80                                         uint32_t clear_value, unsigned flags)
  81 {
  82         struct radeon_winsys_cs *cs = sctx->b.rings.gfx.cs;
  83         uint32_t sync_flag = flags & R600_CP_DMA_SYNC ? S_411_CP_SYNC(1) : 0;
  84         uint32_t raw_wait = flags & SI_CP_DMA_RAW_WAIT ? S_414_RAW_WAIT(1) : 0;
  85         uint32_t dst_sel = flags & CIK_CP_DMA_USE_L2 ? S_411_DSL_SEL(V_411_DST_ADDR_TC_L2) : 0;
  86
  87         assert(size);
  88         assert((size & ((1<<21)-1)) == size);
  89
  90         if (sctx->b.chip_class >= CIK) {
  91                 radeon_emit(cs, PKT3(PKT3_DMA_DATA, 5, 0));
  92                 radeon_emit(cs, sync_flag | dst_sel | S_411_SRC_SEL(V_411_DATA)); /* CP_SYNC [31] | SRC_SEL[30:29] */
  93                 radeon_emit(cs, clear_value);           /* DATA [31:0] */
  94                 radeon_emit(cs, 0);
  95                 radeon_emit(cs, dst_va);                /* DST_ADDR_LO [31:0] */
  96                 radeon_emit(cs, dst_va >> 32);          /* DST_ADDR_HI [15:0] */
  97                 radeon_emit(cs, size | raw_wait);       /* COMMAND [29:22] | BYTE_COUNT [20:0] */
  98         } else {
  99                 radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
 100                 radeon_emit(cs, clear_value);           /* DATA [31:0] */
 101                 radeon_emit(cs, sync_flag | S_411_SRC_SEL(V_411_DATA)); /* CP_SYNC [31] | SRC_SEL[30:29] */
 102                 radeon_emit(cs, dst_va);                        /* DST_ADDR_LO [31:0] */
 103                 radeon_emit(cs, (dst_va >> 32) & 0xffff);       /* DST_ADDR_HI [15:0] */
 104                 radeon_emit(cs, size | raw_wait);               /* COMMAND [29:22] | BYTE_COUNT [20:0] */
 105         }
 106 }
 107
 108 /* The max number of bytes to copy per packet. */
 109 #define CP_DMA_MAX_BYTE_COUNT ((1 << 21) - 8)
 110
 111 static void si_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst,
 112                             unsigned offset, unsigned size, unsigned value,
 113                             bool is_framebuffer)
 114 {
 115         struct si_context *sctx = (struct si_context*)ctx;
 116         unsigned flush_flags, tc_l2_flag;
 117
 118         if (!size)
 119                 return;
 120
 121         /* Mark the buffer range of destination as valid (initialized),
 122          * so that transfer_map knows it should wait for the GPU when mapping
 123          * that range. */
 124         util_range_add(&r600_resource(dst)->valid_buffer_range, offset,
 125                        offset + size);
 126
 127         /* Fallback for unaligned clears. */
 128         if (offset % 4 != 0 || size % 4 != 0) {
 129                 uint32_t *map = sctx->b.ws->buffer_map(r600_resource(dst)->cs_buf,
 130                                                        sctx->b.rings.gfx.cs,
 131                                                        PIPE_TRANSFER_WRITE);
 132                 size /= 4;
 133                 for (unsigned i = 0; i < size; i++)
 134                         *map++ = value;
 135                 return;
 136         }
 137
 138         uint64_t va = r600_resource(dst)->gpu_address + offset;
 139
 140         /* Flush the caches where the resource is bound. */
 141         if (is_framebuffer) {
 142                 flush_flags = SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER;
 143                 tc_l2_flag = 0;
 144         } else {
 145                 flush_flags = SI_CONTEXT_INV_TC_L1 |
 146                               (sctx->b.chip_class == SI ? SI_CONTEXT_INV_TC_L2 : 0) |
 147                               SI_CONTEXT_INV_KCACHE;
 148                 tc_l2_flag = sctx->b.chip_class == SI ? 0 : CIK_CP_DMA_USE_L2;
 149         }
 150
 151         sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
 152                          flush_flags;
 153
 154         while (size) {
 155                 unsigned byte_count = MIN2(size, CP_DMA_MAX_BYTE_COUNT);
 156                 unsigned dma_flags = tc_l2_flag;
 157
 158                 si_need_cs_space(sctx, 7 + (sctx->b.flags ? sctx->cache_flush.num_dw : 0),
 159                                  FALSE);
 160
 161                 /* This must be done after need_cs_space. */
 162                 radeon_add_to_buffer_list(&sctx->b, &sctx->b.rings.gfx,
 163                                       (struct r600_resource*)dst, RADEON_USAGE_WRITE,
 164                                       RADEON_PRIO_MIN);
 165
 166                 /* Flush the caches for the first copy only.
 167                  * Also wait for the previous CP DMA operations. */
 168                 if (sctx->b.flags) {
 169                         si_emit_cache_flush(sctx, NULL);
 170                         dma_flags |= SI_CP_DMA_RAW_WAIT; /* same as WAIT_UNTIL=CP_DMA_IDLE */
 171                 }
 172
 173                 /* Do the synchronization after the last copy, so that all data is written to memory. */
 174                 if (size == byte_count)
 175                         dma_flags |= R600_CP_DMA_SYNC;
 176
 177                 /* Emit the clear packet. */
 178                 si_emit_cp_dma_clear_buffer(sctx, va, byte_count, value, dma_flags);
 179
 180                 size -= byte_count;
 181                 va += byte_count;
 182         }
 183
 184         /* Flush the caches again in case the 3D engine has been prefetching
 185          * the resource. */
 186         sctx->b.flags |= flush_flags;
 187
 188         if (tc_l2_flag)
 189                 r600_resource(dst)->TC_L2_dirty = true;
 190 }
 191
 192 void si_copy_buffer(struct si_context *sctx,
 193                     struct pipe_resource *dst, struct pipe_resource *src,
 194                     uint64_t dst_offset, uint64_t src_offset, unsigned size,
 195                     bool is_framebuffer)
 196 {
 197         unsigned flush_flags, tc_l2_flag;
 198
 199         if (!size)
 200                 return;
 201
 202         /* Mark the buffer range of destination as valid (initialized),
 203          * so that transfer_map knows it should wait for the GPU when mapping
 204          * that range. */
 205         util_range_add(&r600_resource(dst)->valid_buffer_range, dst_offset,
 206                        dst_offset + size);
 207
 208         dst_offset += r600_resource(dst)->gpu_address;
 209         src_offset += r600_resource(src)->gpu_address;
 210
 211         /* Flush the caches where the resource is bound. */
 212         if (is_framebuffer) {
 213                 flush_flags = SI_CONTEXT_FLUSH_AND_INV_FRAMEBUFFER;
 214                 tc_l2_flag = 0;
 215         } else {
 216                 flush_flags = SI_CONTEXT_INV_TC_L1 |
 217                               (sctx->b.chip_class == SI ? SI_CONTEXT_INV_TC_L2 : 0) |
 218                               SI_CONTEXT_INV_KCACHE;
 219                 tc_l2_flag = sctx->b.chip_class == SI ? 0 : CIK_CP_DMA_USE_L2;
 220         }
 221
 222         sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
 223                          flush_flags;
 224
 225         while (size) {
 226                 unsigned sync_flags = tc_l2_flag;
 227                 unsigned byte_count = MIN2(size, CP_DMA_MAX_BYTE_COUNT);
 228
 229                 si_need_cs_space(sctx, 7 + (sctx->b.flags ? sctx->cache_flush.num_dw : 0), FALSE);
 230
 231                 /* Flush the caches for the first copy only. Also wait for old CP DMA packets to complete. */
 232                 if (sctx->b.flags) {
 233                         si_emit_cache_flush(sctx, NULL);
 234                         sync_flags |= SI_CP_DMA_RAW_WAIT;
 235                 }
 236
 237                 /* Do the synchronization after the last copy, so that all data is written to memory. */
 238                 if (size == byte_count) {
 239                         sync_flags |= R600_CP_DMA_SYNC;
 240                 }
 241
 242                 /* This must be done after r600_need_cs_space. */
 243                 radeon_add_to_buffer_list(&sctx->b, &sctx->b.rings.gfx, (struct r600_resource*)src,
 244                                       RADEON_USAGE_READ, RADEON_PRIO_MIN);
 245                 radeon_add_to_buffer_list(&sctx->b, &sctx->b.rings.gfx, (struct r600_resource*)dst,
 246                                       RADEON_USAGE_WRITE, RADEON_PRIO_MIN);
 247
 248                 si_emit_cp_dma_copy_buffer(sctx, dst_offset, src_offset, byte_count, sync_flags);
 249
 250                 size -= byte_count;
 251                 src_offset += byte_count;
 252                 dst_offset += byte_count;
 253         }
 254
 255         /* Flush the caches again in case the 3D engine has been prefetching
 256          * the resource. */
 257         sctx->b.flags |= flush_flags;
 258
 259         if (tc_l2_flag)
 260                 r600_resource(dst)->TC_L2_dirty = true;
 261 }
 262
 263 void si_init_cp_dma_functions(struct si_context *sctx)
 264 {
 265         sctx->b.clear_buffer = si_clear_buffer;
 266 }