void r600_dma_copy_buffer(struct r600_context *rctx,
struct pipe_resource *dst,
struct pipe_resource *src,
uint64_t dst_offset,
uint64_t src_offset,
uint64_t size)
{
struct radeon_winsys_cs *cs = rctx->b.dma.cs;
unsigned i, ncopy, csize;
struct r600_resource *rdst = (struct r600_resource*)dst;
struct r600_resource *rsrc = (struct r600_resource*)src;
/* Mark the buffer range of destination as valid (initialized),
* so that transfer_map knows it should wait for the GPU when mapping
* that range. */
util_range_add(&rdst->valid_buffer_range, dst_offset,
dst_offset + size);
size >>= 2; /* convert to dwords */
ncopy = (size / R600_DMA_COPY_MAX_SIZE_DW) + !!(size % R600_DMA_COPY_MAX_SIZE_DW);
r600_need_dma_space(&rctx->b, ncopy * 5, rdst, rsrc);
for (i = 0; i < ncopy; i++) {
csize = size < R600_DMA_COPY_MAX_SIZE_DW ? size : R600_DMA_COPY_MAX_SIZE_DW;
/* emit reloc before writing cs so that cs is always in consistent state */
radeon_add_to_buffer_list(&rctx->b, &rctx->b.dma, rsrc, RADEON_USAGE_READ,
RADEON_PRIO_SDMA_BUFFER);
radeon_add_to_buffer_list(&rctx->b, &rctx->b.dma, rdst, RADEON_USAGE_WRITE,
RADEON_PRIO_SDMA_BUFFER);
radeon_emit(cs, DMA_PACKET(DMA_PACKET_COPY, 0, 0, csize));
radeon_emit(cs, dst_offset & 0xfffffffc);
radeon_emit(cs, src_offset & 0xfffffffc);
radeon_emit(cs, (dst_offset >> 32UL) & 0xff);
radeon_emit(cs, (src_offset >> 32UL) & 0xff);
dst_offset += csize << 2;
src_offset += csize << 2;
size -= csize;
}
r600_dma_emit_wait_idle(&rctx->b);
}
static void cik_sdma_do_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)
{
struct radeon_winsys_cs *cs = ctx->b.dma.cs;
unsigned i, ncopy, csize;
struct r600_resource *rdst = (struct r600_resource*)dst;
struct r600_resource *rsrc = (struct r600_resource*)src;
dst_offset += r600_resource(dst)->gpu_address;
src_offset += r600_resource(src)->gpu_address;
ncopy = (size + CIK_SDMA_COPY_MAX_SIZE - 1) / CIK_SDMA_COPY_MAX_SIZE;
r600_need_dma_space(&ctx->b, ncopy * 7);
radeon_add_to_buffer_list(&ctx->b, &ctx->b.dma, rsrc, RADEON_USAGE_READ,
RADEON_PRIO_SDMA_BUFFER);
radeon_add_to_buffer_list(&ctx->b, &ctx->b.dma, rdst, RADEON_USAGE_WRITE,
RADEON_PRIO_SDMA_BUFFER);
for (i = 0; i < ncopy; i++) {
csize = size < CIK_SDMA_COPY_MAX_SIZE ? size : CIK_SDMA_COPY_MAX_SIZE;
cs->buf[cs->cdw++] = CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
CIK_SDMA_COPY_SUB_OPCODE_LINEAR,
0);
cs->buf[cs->cdw++] = csize;
cs->buf[cs->cdw++] = 0; /* src/dst endian swap */
cs->buf[cs->cdw++] = src_offset;
cs->buf[cs->cdw++] = src_offset >> 32;
cs->buf[cs->cdw++] = dst_offset;
cs->buf[cs->cdw++] = dst_offset >> 32;
dst_offset += csize;
src_offset += csize;
size -= csize;
}
}
/* This is required to prevent read-after-write hazards. */
void r600_dma_emit_wait_idle(struct r600_common_context *rctx)
{
struct radeon_winsys_cs *cs = rctx->dma.cs;
/* done at the end of DMA calls, so increment this. */
rctx->num_dma_calls++;
/* IBs using too little memory are limited by the IB submission overhead.
* IBs using too much memory are limited by the kernel/TTM overhead.
* Too long IBs create CPU-GPU pipeline bubbles and add latency.
*
* This heuristic makes sure that DMA requests are executed
* very soon after the call is made and lowers memory usage.
* It improves texture upload performance by keeping the DMA
* engine busy while uploads are being submitted.
*/
if (rctx->ws->cs_query_memory_usage(rctx->dma.cs) > 64 * 1024 * 1024) {
rctx->dma.flush(rctx, RADEON_FLUSH_ASYNC, NULL);
return;
}
r600_need_dma_space(rctx, 1, NULL, NULL);
if (!radeon_emitted(cs, 0)) /* empty queue */
return;
/* NOP waits for idle on Evergreen and later. */
if (rctx->chip_class >= CIK)
radeon_emit(cs, 0x00000000); /* NOP */
else if (rctx->chip_class >= EVERGREEN)
radeon_emit(cs, 0xf0000000); /* NOP */
else {
/* TODO: R600-R700 should use the FENCE packet.
* CS checker support is required. */
}
}
static void cik_sdma_copy_tile(struct si_context *ctx,
struct pipe_resource *dst,
unsigned dst_level,
struct pipe_resource *src,
unsigned src_level,
unsigned y,
unsigned copy_height,
unsigned y_align,
unsigned pitch,
unsigned bpe)
{
struct radeon_winsys_cs *cs = ctx->b.dma.cs;
struct si_screen *sscreen = ctx->screen;
struct r600_texture *rsrc = (struct r600_texture*)src;
struct r600_texture *rdst = (struct r600_texture*)dst;
struct r600_texture *rlinear, *rtiled;
unsigned linear_lvl, tiled_lvl;
unsigned array_mode, lbpe, pitch_tile_max, slice_tile_max, size;
unsigned ncopy, height, cheight, detile, i, src_mode, dst_mode;
unsigned sub_op, bank_h, bank_w, mt_aspect, nbanks, tile_split, mt;
uint64_t base, addr;
unsigned pipe_config, tile_mode_index;
dst_mode = rdst->surface.level[dst_level].mode;
src_mode = rsrc->surface.level[src_level].mode;
assert(dst_mode != src_mode);
assert(src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED || dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED);
sub_op = CIK_SDMA_COPY_SUB_OPCODE_TILED;
lbpe = util_logbase2(bpe);
pitch_tile_max = ((pitch / bpe) / 8) - 1;
detile = dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED;
rlinear = detile ? rdst : rsrc;
rtiled = detile ? rsrc : rdst;
linear_lvl = detile ? dst_level : src_level;
tiled_lvl = detile ? src_level : dst_level;
assert(!util_format_is_depth_and_stencil(rtiled->resource.b.b.format));
array_mode = si_array_mode(rtiled->surface.level[tiled_lvl].mode);
slice_tile_max = (rtiled->surface.level[tiled_lvl].nblk_x *
rtiled->surface.level[tiled_lvl].nblk_y) / (8*8) - 1;
height = rlinear->surface.level[linear_lvl].nblk_y;
base = rtiled->surface.level[tiled_lvl].offset;
addr = rlinear->surface.level[linear_lvl].offset;
bank_h = cik_bank_wh(rtiled->surface.bankh);
bank_w = cik_bank_wh(rtiled->surface.bankw);
mt_aspect = cik_macro_tile_aspect(rtiled->surface.mtilea);
tile_split = cik_tile_split(rtiled->surface.tile_split);
tile_mode_index = si_tile_mode_index(rtiled, tiled_lvl, false);
nbanks = si_num_banks(sscreen, rtiled);
base += rtiled->resource.gpu_address;
addr += rlinear->resource.gpu_address;
pipe_config = cik_db_pipe_config(sscreen, tile_mode_index);
mt = cik_micro_tile_mode(sscreen, tile_mode_index);
size = (copy_height * pitch) / 4;
cheight = copy_height;
if (((cheight * pitch) / 4) > CIK_SDMA_COPY_MAX_SIZE) {
cheight = (CIK_SDMA_COPY_MAX_SIZE * 4) / pitch;
cheight &= ~(y_align - 1);
}
ncopy = (copy_height + cheight - 1) / cheight;
r600_need_dma_space(&ctx->b, ncopy * 12);
radeon_add_to_buffer_list(&ctx->b, &ctx->b.dma, &rsrc->resource,
RADEON_USAGE_READ, RADEON_PRIO_SDMA_TEXTURE);
radeon_add_to_buffer_list(&ctx->b, &ctx->b.dma, &rdst->resource,
RADEON_USAGE_WRITE, RADEON_PRIO_SDMA_TEXTURE);
copy_height = size * 4 / pitch;
for (i = 0; i < ncopy; i++) {
cheight = copy_height;
if (((cheight * pitch) / 4) > CIK_SDMA_COPY_MAX_SIZE) {
cheight = (CIK_SDMA_COPY_MAX_SIZE * 4) / pitch;
cheight &= ~(y_align - 1);
}
size = (cheight * pitch) / 4;
cs->buf[cs->cdw++] = CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
sub_op, detile << 15);
cs->buf[cs->cdw++] = base;
cs->buf[cs->cdw++] = base >> 32;
cs->buf[cs->cdw++] = ((height - 1) << 16) | pitch_tile_max;
cs->buf[cs->cdw++] = slice_tile_max;
cs->buf[cs->cdw++] = (pipe_config << 26) | (mt_aspect << 24) |
(nbanks << 21) | (bank_h << 18) | (bank_w << 15) |
(tile_split << 11) | (mt << 8) | (array_mode << 3) |
lbpe;
cs->buf[cs->cdw++] = y << 16; /* | x */
cs->buf[cs->cdw++] = 0; /* z */
cs->buf[cs->cdw++] = addr & 0xfffffffc;
cs->buf[cs->cdw++] = addr >> 32;
cs->buf[cs->cdw++] = (pitch / bpe) - 1;
cs->buf[cs->cdw++] = size;
copy_height -= cheight;
y += cheight;
}
}
