/*
* Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
- * Copyright 2014,2015 Advanced Micro Devices, Inc.
+ * Copyright 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
* OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
* USE OR OTHER DEALINGS IN THE SOFTWARE.
- *
- * Authors:
- * Jerome Glisse
*/
-#include "sid.h"
#include "si_pipe.h"
-#include "radeon/r600_cs.h"
-
-#include "util/u_format.h"
+#include "sid.h"
-static uint32_t cik_micro_tile_mode(struct si_screen *sscreen, unsigned tile_mode)
+static unsigned minify_as_blocks(unsigned width, unsigned level, unsigned blk_w)
{
- if (sscreen->b.info.si_tile_mode_array_valid) {
- uint32_t gb_tile_mode = sscreen->b.info.si_tile_mode_array[tile_mode];
-
- return G_009910_MICRO_TILE_MODE_NEW(gb_tile_mode);
- }
-
- /* The kernel cannod return the tile mode array. Guess? */
- return V_009910_ADDR_SURF_THIN_MICRO_TILING;
+ width = u_minify(width, level);
+ return DIV_ROUND_UP(width, blk_w);
}
-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)
+static unsigned encode_tile_info(struct si_context *sctx, struct si_texture *tex, unsigned level,
+ bool set_bpp)
{
- 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;
- }
+ struct radeon_info *info = &sctx->screen->info;
+ unsigned tile_index = tex->surface.u.legacy.tiling_index[level];
+ unsigned macro_tile_index = tex->surface.u.legacy.macro_tile_index;
+ unsigned tile_mode = info->si_tile_mode_array[tile_index];
+ unsigned macro_tile_mode = info->cik_macrotile_mode_array[macro_tile_index];
+
+ return (set_bpp ? util_logbase2(tex->surface.bpe) : 0) | (G_009910_ARRAY_MODE(tile_mode) << 3) |
+ (G_009910_MICRO_TILE_MODE_NEW(tile_mode) << 8) |
+ /* Non-depth modes don't have TILE_SPLIT set. */
+ ((util_logbase2(tex->surface.u.legacy.tile_split >> 6)) << 11) |
+ (G_009990_BANK_WIDTH(macro_tile_mode) << 15) |
+ (G_009990_BANK_HEIGHT(macro_tile_mode) << 18) |
+ (G_009990_NUM_BANKS(macro_tile_mode) << 21) |
+ (G_009990_MACRO_TILE_ASPECT(macro_tile_mode) << 24) |
+ (G_009910_PIPE_CONFIG(tile_mode) << 26);
}
-static void cik_sdma_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)
+static bool si_sdma_v4_copy_texture(struct si_context *sctx, struct pipe_resource *dst,
+ unsigned dst_level, unsigned dstx, unsigned dsty, unsigned dstz,
+ struct pipe_resource *src, unsigned src_level,
+ const struct pipe_box *src_box)
{
- struct r600_resource *rdst = (struct r600_resource*)dst;
-
- /* 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);
-
- cik_sdma_do_copy_buffer(ctx, dst, src, dst_offset, src_offset, size);
+ struct si_texture *ssrc = (struct si_texture *)src;
+ struct si_texture *sdst = (struct si_texture *)dst;
+
+ unsigned bpp = sdst->surface.bpe;
+ uint64_t dst_address = sdst->buffer.gpu_address + sdst->surface.u.gfx9.surf_offset;
+ uint64_t src_address = ssrc->buffer.gpu_address + ssrc->surface.u.gfx9.surf_offset;
+ unsigned dst_pitch = sdst->surface.u.gfx9.surf_pitch;
+ unsigned src_pitch = ssrc->surface.u.gfx9.surf_pitch;
+ uint64_t dst_slice_pitch = ((uint64_t)sdst->surface.u.gfx9.surf_slice_size) / bpp;
+ uint64_t src_slice_pitch = ((uint64_t)ssrc->surface.u.gfx9.surf_slice_size) / bpp;
+ unsigned srcx = src_box->x / ssrc->surface.blk_w;
+ unsigned srcy = src_box->y / ssrc->surface.blk_h;
+ unsigned srcz = src_box->z;
+ unsigned copy_width = DIV_ROUND_UP(src_box->width, ssrc->surface.blk_w);
+ unsigned copy_height = DIV_ROUND_UP(src_box->height, ssrc->surface.blk_h);
+ unsigned copy_depth = src_box->depth;
+ unsigned xalign = MAX2(1, 4 / bpp);
+
+ assert(src_level <= src->last_level);
+ assert(dst_level <= dst->last_level);
+ assert(sdst->surface.u.gfx9.surf_offset + dst_slice_pitch * bpp * (dstz + src_box->depth) <=
+ sdst->buffer.buf->size);
+ assert(ssrc->surface.u.gfx9.surf_offset + src_slice_pitch * bpp * (srcz + src_box->depth) <=
+ ssrc->buffer.buf->size);
+
+ if (!si_prepare_for_dma_blit(sctx, sdst, dst_level, dstx, dsty, dstz, ssrc, src_level, src_box))
+ return false;
+
+ dstx /= sdst->surface.blk_w;
+ dsty /= sdst->surface.blk_h;
+
+ if (srcx >= (1 << 14) || srcy >= (1 << 14) || srcz >= (1 << 11) || dstx >= (1 << 14) ||
+ dsty >= (1 << 14) || dstz >= (1 << 11))
+ return false;
+
+ /* Linear -> linear sub-window copy. */
+ if (ssrc->surface.is_linear && sdst->surface.is_linear) {
+ struct radeon_cmdbuf *cs = sctx->sdma_cs;
+
+ /* Check if everything fits into the bitfields */
+ if (!(src_pitch <= (1 << 19) && dst_pitch <= (1 << 19) && src_slice_pitch <= (1 << 28) &&
+ dst_slice_pitch <= (1 << 28) && copy_width <= (1 << 14) && copy_height <= (1 << 14) &&
+ copy_depth <= (1 << 11)))
+ return false;
+
+ si_need_dma_space(sctx, 13, &sdst->buffer, &ssrc->buffer);
+
+ src_address += ssrc->surface.u.gfx9.offset[src_level];
+ dst_address += sdst->surface.u.gfx9.offset[dst_level];
+
+ /* Check alignments */
+ if ((src_address % 4) != 0 || (dst_address % 4) != 0 || (src_pitch % xalign) != 0)
+ return false;
+
+ radeon_emit(
+ cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY, CIK_SDMA_COPY_SUB_OPCODE_LINEAR_SUB_WINDOW,
+ sctx->ws->cs_is_secure(cs) ? (1u << 2) : 0) |
+ (util_logbase2(bpp) << 29));
+ radeon_emit(cs, src_address);
+ radeon_emit(cs, src_address >> 32);
+ radeon_emit(cs, srcx | (srcy << 16));
+ radeon_emit(cs, srcz | ((src_pitch - 1) << 13));
+ radeon_emit(cs, src_slice_pitch - 1);
+ radeon_emit(cs, dst_address);
+ radeon_emit(cs, dst_address >> 32);
+ radeon_emit(cs, dstx | (dsty << 16));
+ radeon_emit(cs, dstz | ((dst_pitch - 1) << 13));
+ radeon_emit(cs, dst_slice_pitch - 1);
+ radeon_emit(cs, (copy_width - 1) | ((copy_height - 1) << 16));
+ radeon_emit(cs, (copy_depth - 1));
+ return true;
+ }
+
+ /* Linear <-> Tiled sub-window copy */
+ if (ssrc->surface.is_linear != sdst->surface.is_linear) {
+ struct si_texture *tiled = ssrc->surface.is_linear ? sdst : ssrc;
+ struct si_texture *linear = tiled == ssrc ? sdst : ssrc;
+ unsigned tiled_level = tiled == ssrc ? src_level : dst_level;
+ unsigned linear_level = linear == ssrc ? src_level : dst_level;
+ unsigned tiled_x = tiled == ssrc ? srcx : dstx;
+ unsigned linear_x = linear == ssrc ? srcx : dstx;
+ unsigned tiled_y = tiled == ssrc ? srcy : dsty;
+ unsigned linear_y = linear == ssrc ? srcy : dsty;
+ unsigned tiled_z = tiled == ssrc ? srcz : dstz;
+ unsigned linear_z = linear == ssrc ? srcz : dstz;
+ unsigned tiled_width = tiled == ssrc
+ ? DIV_ROUND_UP(ssrc->buffer.b.b.width0, ssrc->surface.blk_w)
+ : DIV_ROUND_UP(sdst->buffer.b.b.width0, sdst->surface.blk_w);
+ unsigned tiled_height = tiled == ssrc
+ ? DIV_ROUND_UP(ssrc->buffer.b.b.height0, ssrc->surface.blk_h)
+ : DIV_ROUND_UP(sdst->buffer.b.b.height0, sdst->surface.blk_h);
+ unsigned tiled_depth = tiled == ssrc ? ssrc->buffer.b.b.depth0 : sdst->buffer.b.b.depth0;
+ unsigned linear_pitch = linear == ssrc ? src_pitch : dst_pitch;
+ unsigned linear_slice_pitch = linear == ssrc ? src_slice_pitch : dst_slice_pitch;
+ uint64_t tiled_address = tiled == ssrc ? src_address : dst_address;
+ uint64_t linear_address = linear == ssrc ? src_address : dst_address;
+ struct radeon_cmdbuf *cs = sctx->sdma_cs;
+
+ linear_address += linear->surface.u.gfx9.offset[linear_level];
+
+ /* Check if everything fits into the bitfields */
+ if (!(tiled_x <= (1 << 14) && tiled_y <= (1 << 14) && tiled_z <= (1 << 11) &&
+ tiled_width <= (1 << 14) && tiled_height <= (1 << 14) && tiled_depth <= (1 << 11) &&
+ linear_x <= (1 << 14) && linear_y <= (1 << 14) && linear_z <= (1 << 11) &&
+ linear_pitch <= (1 << 14) && linear_slice_pitch <= (1 << 28) &&
+ copy_width <= (1 << 14) && copy_height <= (1 << 14) && copy_depth <= (1 << 11)))
+ return false;
+
+ /* Check alignments */
+ if ((tiled_address % 256 != 0) || (linear_address % 4 != 0) || (linear_pitch % xalign != 0) ||
+ (linear_slice_pitch % xalign != 0))
+ return false;
+
+ si_need_dma_space(sctx, 14, &sdst->buffer, &ssrc->buffer);
+
+ radeon_emit(
+ cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY, CIK_SDMA_COPY_SUB_OPCODE_TILED_SUB_WINDOW,
+ sctx->ws->cs_is_secure(cs) ? (1u << 2) : 0) |
+ tiled->buffer.b.b.last_level << 20 | tiled_level << 24 |
+ (linear == sdst ? 1u : 0) << 31);
+ radeon_emit(cs, (uint32_t)tiled_address);
+ radeon_emit(cs, (uint32_t)(tiled_address >> 32));
+ radeon_emit(cs, tiled_x | (tiled_y << 16));
+ radeon_emit(cs, tiled_z | ((tiled_width - 1) << 16));
+ radeon_emit(cs, (tiled_height - 1) | (tiled_depth - 1) << 16);
+ radeon_emit(cs, util_logbase2(bpp) | tiled->surface.u.gfx9.surf.swizzle_mode << 3 |
+ tiled->surface.u.gfx9.resource_type << 9 |
+ tiled->surface.u.gfx9.surf.epitch << 16);
+ radeon_emit(cs, (uint32_t)linear_address);
+ radeon_emit(cs, (uint32_t)(linear_address >> 32));
+ radeon_emit(cs, linear_x | (linear_y << 16));
+ radeon_emit(cs, linear_z | ((linear_pitch - 1) << 16));
+ radeon_emit(cs, linear_slice_pitch - 1);
+ radeon_emit(cs, (copy_width - 1) | ((copy_height - 1) << 16));
+ radeon_emit(cs, (copy_depth - 1));
+ return true;
+ }
+
+ return false;
}
-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)
+static bool cik_sdma_copy_texture(struct si_context *sctx, struct pipe_resource *dst,
+ unsigned dst_level, unsigned dstx, unsigned dsty, unsigned dstz,
+ struct pipe_resource *src, unsigned src_level,
+ const struct pipe_box *src_box)
{
- 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;
- /* downcast linear aligned to linear to simplify test */
- src_mode = src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED ? RADEON_SURF_MODE_LINEAR : src_mode;
- dst_mode = dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED ? RADEON_SURF_MODE_LINEAR : dst_mode;
- assert(dst_mode != src_mode);
- assert(src_mode == RADEON_SURF_MODE_LINEAR || dst_mode == RADEON_SURF_MODE_LINEAR);
-
- 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;
- 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;
- }
+ struct radeon_info *info = &sctx->screen->info;
+ struct si_texture *ssrc = (struct si_texture *)src;
+ struct si_texture *sdst = (struct si_texture *)dst;
+ unsigned bpp = sdst->surface.bpe;
+ uint64_t dst_address = sdst->buffer.gpu_address + sdst->surface.u.legacy.level[dst_level].offset;
+ uint64_t src_address = ssrc->buffer.gpu_address + ssrc->surface.u.legacy.level[src_level].offset;
+ unsigned dst_mode = sdst->surface.u.legacy.level[dst_level].mode;
+ unsigned src_mode = ssrc->surface.u.legacy.level[src_level].mode;
+ unsigned dst_tile_index = sdst->surface.u.legacy.tiling_index[dst_level];
+ unsigned src_tile_index = ssrc->surface.u.legacy.tiling_index[src_level];
+ unsigned dst_tile_mode = info->si_tile_mode_array[dst_tile_index];
+ unsigned src_tile_mode = info->si_tile_mode_array[src_tile_index];
+ unsigned dst_micro_mode = G_009910_MICRO_TILE_MODE_NEW(dst_tile_mode);
+ unsigned src_micro_mode = G_009910_MICRO_TILE_MODE_NEW(src_tile_mode);
+ unsigned dst_tile_swizzle = dst_mode == RADEON_SURF_MODE_2D ? sdst->surface.tile_swizzle : 0;
+ unsigned src_tile_swizzle = src_mode == RADEON_SURF_MODE_2D ? ssrc->surface.tile_swizzle : 0;
+ unsigned dst_pitch = sdst->surface.u.legacy.level[dst_level].nblk_x;
+ unsigned src_pitch = ssrc->surface.u.legacy.level[src_level].nblk_x;
+ uint64_t dst_slice_pitch =
+ ((uint64_t)sdst->surface.u.legacy.level[dst_level].slice_size_dw * 4) / bpp;
+ uint64_t src_slice_pitch =
+ ((uint64_t)ssrc->surface.u.legacy.level[src_level].slice_size_dw * 4) / bpp;
+ unsigned dst_width = minify_as_blocks(sdst->buffer.b.b.width0, dst_level, sdst->surface.blk_w);
+ unsigned src_width = minify_as_blocks(ssrc->buffer.b.b.width0, src_level, ssrc->surface.blk_w);
+ unsigned dst_height = minify_as_blocks(sdst->buffer.b.b.height0, dst_level, sdst->surface.blk_h);
+ unsigned src_height = minify_as_blocks(ssrc->buffer.b.b.height0, src_level, ssrc->surface.blk_h);
+ unsigned srcx = src_box->x / ssrc->surface.blk_w;
+ unsigned srcy = src_box->y / ssrc->surface.blk_h;
+ unsigned srcz = src_box->z;
+ unsigned copy_width = DIV_ROUND_UP(src_box->width, ssrc->surface.blk_w);
+ unsigned copy_height = DIV_ROUND_UP(src_box->height, ssrc->surface.blk_h);
+ unsigned copy_depth = src_box->depth;
+
+ assert(src_level <= src->last_level);
+ assert(dst_level <= dst->last_level);
+ assert(sdst->surface.u.legacy.level[dst_level].offset +
+ dst_slice_pitch * bpp * (dstz + src_box->depth) <=
+ sdst->buffer.buf->size);
+ assert(ssrc->surface.u.legacy.level[src_level].offset +
+ src_slice_pitch * bpp * (srcz + src_box->depth) <=
+ ssrc->buffer.buf->size);
+
+ if (!si_prepare_for_dma_blit(sctx, sdst, dst_level, dstx, dsty, dstz, ssrc, src_level, src_box))
+ return false;
+
+ dstx /= sdst->surface.blk_w;
+ dsty /= sdst->surface.blk_h;
+
+ if (srcx >= (1 << 14) || srcy >= (1 << 14) || srcz >= (1 << 11) || dstx >= (1 << 14) ||
+ dsty >= (1 << 14) || dstz >= (1 << 11))
+ return false;
+
+ dst_address |= dst_tile_swizzle << 8;
+ src_address |= src_tile_swizzle << 8;
+
+ /* Linear -> linear sub-window copy. */
+ if (dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED && src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED &&
+ /* check if everything fits into the bitfields */
+ src_pitch <= (1 << 14) && dst_pitch <= (1 << 14) && src_slice_pitch <= (1 << 28) &&
+ dst_slice_pitch <= (1 << 28) && copy_width <= (1 << 14) && copy_height <= (1 << 14) &&
+ copy_depth <= (1 << 11) &&
+ /* HW limitation - GFX7: */
+ (sctx->chip_class != GFX7 ||
+ (copy_width < (1 << 14) && copy_height < (1 << 14) && copy_depth < (1 << 11))) &&
+ /* HW limitation - some GFX7 parts: */
+ ((sctx->family != CHIP_BONAIRE && sctx->family != CHIP_KAVERI) ||
+ (srcx + copy_width != (1 << 14) && srcy + copy_height != (1 << 14)))) {
+ struct radeon_cmdbuf *cs = sctx->sdma_cs;
+
+ si_need_dma_space(sctx, 13, &sdst->buffer, &ssrc->buffer);
+
+ radeon_emit(
+ cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY, CIK_SDMA_COPY_SUB_OPCODE_LINEAR_SUB_WINDOW, 0) |
+ (util_logbase2(bpp) << 29));
+ radeon_emit(cs, src_address);
+ radeon_emit(cs, src_address >> 32);
+ radeon_emit(cs, srcx | (srcy << 16));
+ radeon_emit(cs, srcz | ((src_pitch - 1) << 16));
+ radeon_emit(cs, src_slice_pitch - 1);
+ radeon_emit(cs, dst_address);
+ radeon_emit(cs, dst_address >> 32);
+ radeon_emit(cs, dstx | (dsty << 16));
+ radeon_emit(cs, dstz | ((dst_pitch - 1) << 16));
+ radeon_emit(cs, dst_slice_pitch - 1);
+ if (sctx->chip_class == GFX7) {
+ radeon_emit(cs, copy_width | (copy_height << 16));
+ radeon_emit(cs, copy_depth);
+ } else {
+ radeon_emit(cs, (copy_width - 1) | ((copy_height - 1) << 16));
+ radeon_emit(cs, (copy_depth - 1));
+ }
+ return true;
+ }
+
+ /* Tiled <-> linear sub-window copy. */
+ if ((src_mode >= RADEON_SURF_MODE_1D) != (dst_mode >= RADEON_SURF_MODE_1D)) {
+ struct si_texture *tiled = src_mode >= RADEON_SURF_MODE_1D ? ssrc : sdst;
+ struct si_texture *linear = tiled == ssrc ? sdst : ssrc;
+ unsigned tiled_level = tiled == ssrc ? src_level : dst_level;
+ unsigned linear_level = linear == ssrc ? src_level : dst_level;
+ unsigned tiled_x = tiled == ssrc ? srcx : dstx;
+ unsigned linear_x = linear == ssrc ? srcx : dstx;
+ unsigned tiled_y = tiled == ssrc ? srcy : dsty;
+ unsigned linear_y = linear == ssrc ? srcy : dsty;
+ unsigned tiled_z = tiled == ssrc ? srcz : dstz;
+ unsigned linear_z = linear == ssrc ? srcz : dstz;
+ unsigned tiled_width = tiled == ssrc ? src_width : dst_width;
+ unsigned linear_width = linear == ssrc ? src_width : dst_width;
+ unsigned tiled_pitch = tiled == ssrc ? src_pitch : dst_pitch;
+ unsigned linear_pitch = linear == ssrc ? src_pitch : dst_pitch;
+ unsigned tiled_slice_pitch = tiled == ssrc ? src_slice_pitch : dst_slice_pitch;
+ unsigned linear_slice_pitch = linear == ssrc ? src_slice_pitch : dst_slice_pitch;
+ uint64_t tiled_address = tiled == ssrc ? src_address : dst_address;
+ uint64_t linear_address = linear == ssrc ? src_address : dst_address;
+ unsigned tiled_micro_mode = tiled == ssrc ? src_micro_mode : dst_micro_mode;
+
+ assert(tiled_pitch % 8 == 0);
+ assert(tiled_slice_pitch % 64 == 0);
+ unsigned pitch_tile_max = tiled_pitch / 8 - 1;
+ unsigned slice_tile_max = tiled_slice_pitch / 64 - 1;
+ unsigned xalign = MAX2(1, 4 / bpp);
+ unsigned copy_width_aligned = copy_width;
+
+ /* If the region ends at the last pixel and is unaligned, we
+ * can copy the remainder of the line that is not visible to
+ * make it aligned.
+ */
+ if (copy_width % xalign != 0 && linear_x + copy_width == linear_width &&
+ tiled_x + copy_width == tiled_width &&
+ linear_x + align(copy_width, xalign) <= linear_pitch &&
+ tiled_x + align(copy_width, xalign) <= tiled_pitch)
+ copy_width_aligned = align(copy_width, xalign);
+
+ /* HW limitations. */
+ if ((sctx->family == CHIP_BONAIRE || sctx->family == CHIP_KAVERI) &&
+ linear_pitch - 1 == 0x3fff && bpp == 16)
+ return false;
+
+ if (sctx->chip_class == GFX7 &&
+ (copy_width_aligned == (1 << 14) || copy_height == (1 << 14) || copy_depth == (1 << 11)))
+ return false;
+
+ if ((sctx->family == CHIP_BONAIRE || sctx->family == CHIP_KAVERI ||
+ sctx->family == CHIP_KABINI) &&
+ (tiled_x + copy_width == (1 << 14) || tiled_y + copy_height == (1 << 14)))
+ return false;
+
+ /* The hw can read outside of the given linear buffer bounds,
+ * or access those pages but not touch the memory in case
+ * of writes. (it still causes a VM fault)
+ *
+ * Out-of-bounds memory access or page directory access must
+ * be prevented.
+ */
+ int64_t start_linear_address, end_linear_address;
+ unsigned granularity;
+
+ /* Deduce the size of reads from the linear surface. */
+ switch (tiled_micro_mode) {
+ case V_009910_ADDR_SURF_DISPLAY_MICRO_TILING:
+ granularity = bpp == 1 ? 64 / (8 * bpp) : 128 / (8 * bpp);
+ break;
+ case V_009910_ADDR_SURF_THIN_MICRO_TILING:
+ case V_009910_ADDR_SURF_DEPTH_MICRO_TILING:
+ if (0 /* TODO: THICK microtiling */)
+ granularity =
+ bpp == 1 ? 32 / (8 * bpp)
+ : bpp == 2 ? 64 / (8 * bpp) : bpp <= 8 ? 128 / (8 * bpp) : 256 / (8 * bpp);
+ else
+ granularity = bpp <= 2 ? 64 / (8 * bpp) : bpp <= 8 ? 128 / (8 * bpp) : 256 / (8 * bpp);
+ break;
+ default:
+ return false;
+ }
+
+ /* The linear reads start at tiled_x & ~(granularity - 1).
+ * If linear_x == 0 && tiled_x % granularity != 0, the hw
+ * starts reading from an address preceding linear_address!!!
+ */
+ start_linear_address =
+ linear->surface.u.legacy.level[linear_level].offset +
+ bpp * (linear_z * linear_slice_pitch + linear_y * linear_pitch + linear_x);
+ start_linear_address -= (int)(bpp * (tiled_x % granularity));
+
+ end_linear_address =
+ linear->surface.u.legacy.level[linear_level].offset +
+ bpp * ((linear_z + copy_depth - 1) * linear_slice_pitch +
+ (linear_y + copy_height - 1) * linear_pitch + (linear_x + copy_width));
+
+ if ((tiled_x + copy_width) % granularity)
+ end_linear_address += granularity - (tiled_x + copy_width) % granularity;
+
+ if (start_linear_address < 0 || end_linear_address > linear->surface.surf_size)
+ return false;
+
+ /* Check requirements. */
+ if (tiled_address % 256 == 0 && linear_address % 4 == 0 && linear_pitch % xalign == 0 &&
+ linear_x % xalign == 0 && tiled_x % xalign == 0 && copy_width_aligned % xalign == 0 &&
+ tiled_micro_mode != V_009910_ADDR_SURF_ROTATED_MICRO_TILING &&
+ /* check if everything fits into the bitfields */
+ tiled->surface.u.legacy.tile_split <= 4096 && pitch_tile_max < (1 << 11) &&
+ slice_tile_max < (1 << 22) && linear_pitch <= (1 << 14) &&
+ linear_slice_pitch <= (1 << 28) && copy_width_aligned <= (1 << 14) &&
+ copy_height <= (1 << 14) && copy_depth <= (1 << 11)) {
+ struct radeon_cmdbuf *cs = sctx->sdma_cs;
+ uint32_t direction = linear == sdst ? 1u << 31 : 0;
+
+ si_need_dma_space(sctx, 14, &sdst->buffer, &ssrc->buffer);
+
+ radeon_emit(cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY,
+ CIK_SDMA_COPY_SUB_OPCODE_TILED_SUB_WINDOW, 0) |
+ direction);
+ radeon_emit(cs, tiled_address);
+ radeon_emit(cs, tiled_address >> 32);
+ radeon_emit(cs, tiled_x | (tiled_y << 16));
+ radeon_emit(cs, tiled_z | (pitch_tile_max << 16));
+ radeon_emit(cs, slice_tile_max);
+ radeon_emit(cs, encode_tile_info(sctx, tiled, tiled_level, true));
+ radeon_emit(cs, linear_address);
+ radeon_emit(cs, linear_address >> 32);
+ radeon_emit(cs, linear_x | (linear_y << 16));
+ radeon_emit(cs, linear_z | ((linear_pitch - 1) << 16));
+ radeon_emit(cs, linear_slice_pitch - 1);
+ if (sctx->chip_class == GFX7) {
+ radeon_emit(cs, copy_width_aligned | (copy_height << 16));
+ radeon_emit(cs, copy_depth);
+ } else {
+ radeon_emit(cs, (copy_width_aligned - 1) | ((copy_height - 1) << 16));
+ radeon_emit(cs, (copy_depth - 1));
+ }
+ return true;
+ }
+ }
+
+ /* Tiled -> Tiled sub-window copy. */
+ if (dst_mode >= RADEON_SURF_MODE_1D && src_mode >= RADEON_SURF_MODE_1D &&
+ /* check if these fit into the bitfields */
+ src_address % 256 == 0 && dst_address % 256 == 0 &&
+ ssrc->surface.u.legacy.tile_split <= 4096 && sdst->surface.u.legacy.tile_split <= 4096 &&
+ dstx % 8 == 0 && dsty % 8 == 0 && srcx % 8 == 0 && srcy % 8 == 0 &&
+ /* this can either be equal, or display->rotated (GFX8+ only) */
+ (src_micro_mode == dst_micro_mode ||
+ (sctx->chip_class >= GFX8 && src_micro_mode == V_009910_ADDR_SURF_DISPLAY_MICRO_TILING &&
+ dst_micro_mode == V_009910_ADDR_SURF_ROTATED_MICRO_TILING))) {
+ assert(src_pitch % 8 == 0);
+ assert(dst_pitch % 8 == 0);
+ assert(src_slice_pitch % 64 == 0);
+ assert(dst_slice_pitch % 64 == 0);
+ unsigned src_pitch_tile_max = src_pitch / 8 - 1;
+ unsigned dst_pitch_tile_max = dst_pitch / 8 - 1;
+ unsigned src_slice_tile_max = src_slice_pitch / 64 - 1;
+ unsigned dst_slice_tile_max = dst_slice_pitch / 64 - 1;
+ unsigned copy_width_aligned = copy_width;
+ unsigned copy_height_aligned = copy_height;
+
+ /* If the region ends at the last pixel and is unaligned, we
+ * can copy the remainder of the tile that is not visible to
+ * make it aligned.
+ */
+ if (copy_width % 8 != 0 && srcx + copy_width == src_width && dstx + copy_width == dst_width)
+ copy_width_aligned = align(copy_width, 8);
+
+ if (copy_height % 8 != 0 && srcy + copy_height == src_height &&
+ dsty + copy_height == dst_height)
+ copy_height_aligned = align(copy_height, 8);
+
+ /* check if these fit into the bitfields */
+ if (src_pitch_tile_max < (1 << 11) && dst_pitch_tile_max < (1 << 11) &&
+ src_slice_tile_max < (1 << 22) && dst_slice_tile_max < (1 << 22) &&
+ copy_width_aligned <= (1 << 14) && copy_height_aligned <= (1 << 14) &&
+ copy_depth <= (1 << 11) && copy_width_aligned % 8 == 0 && copy_height_aligned % 8 == 0 &&
+ /* HW limitation - GFX7: */
+ (sctx->chip_class != GFX7 ||
+ (copy_width_aligned < (1 << 14) && copy_height_aligned < (1 << 14) &&
+ copy_depth < (1 << 11))) &&
+ /* HW limitation - some GFX7 parts: */
+ ((sctx->family != CHIP_BONAIRE && sctx->family != CHIP_KAVERI &&
+ sctx->family != CHIP_KABINI) ||
+ (srcx + copy_width_aligned != (1 << 14) && srcy + copy_height_aligned != (1 << 14) &&
+ dstx + copy_width != (1 << 14)))) {
+ struct radeon_cmdbuf *cs = sctx->sdma_cs;
+
+ si_need_dma_space(sctx, 15, &sdst->buffer, &ssrc->buffer);
+
+ radeon_emit(
+ cs, CIK_SDMA_PACKET(CIK_SDMA_OPCODE_COPY, CIK_SDMA_COPY_SUB_OPCODE_T2T_SUB_WINDOW, 0));
+ radeon_emit(cs, src_address);
+ radeon_emit(cs, src_address >> 32);
+ radeon_emit(cs, srcx | (srcy << 16));
+ radeon_emit(cs, srcz | (src_pitch_tile_max << 16));
+ radeon_emit(cs, src_slice_tile_max);
+ radeon_emit(cs, encode_tile_info(sctx, ssrc, src_level, true));
+ radeon_emit(cs, dst_address);
+ radeon_emit(cs, dst_address >> 32);
+ radeon_emit(cs, dstx | (dsty << 16));
+ radeon_emit(cs, dstz | (dst_pitch_tile_max << 16));
+ radeon_emit(cs, dst_slice_tile_max);
+ radeon_emit(cs, encode_tile_info(sctx, sdst, dst_level, false));
+ if (sctx->chip_class == GFX7) {
+ radeon_emit(cs, copy_width_aligned | (copy_height_aligned << 16));
+ radeon_emit(cs, copy_depth);
+ } else {
+ radeon_emit(cs, (copy_width_aligned - 8) | ((copy_height_aligned - 8) << 16));
+ radeon_emit(cs, (copy_depth - 1));
+ }
+ return true;
+ }
+ }
+
+ return false;
}
-void cik_sdma_copy(struct pipe_context *ctx,
- struct pipe_resource *dst,
- unsigned dst_level,
- unsigned dstx, unsigned dsty, unsigned dstz,
- struct pipe_resource *src,
- unsigned src_level,
- const struct pipe_box *src_box)
+static void cik_sdma_copy(struct pipe_context *ctx, struct pipe_resource *dst, unsigned dst_level,
+ unsigned dstx, unsigned dsty, unsigned dstz, struct pipe_resource *src,
+ unsigned src_level, const struct pipe_box *src_box)
{
- struct si_context *sctx = (struct si_context *)ctx;
- struct r600_texture *rsrc = (struct r600_texture*)src;
- struct r600_texture *rdst = (struct r600_texture*)dst;
- unsigned dst_pitch, src_pitch, bpe, dst_mode, src_mode;
- unsigned src_w, dst_w;
- unsigned src_x, src_y;
- unsigned copy_height, y_align;
- unsigned dst_x = dstx, dst_y = dsty, dst_z = dstz;
-
- if (sctx->b.dma.cs == NULL) {
- goto fallback;
- }
-
- if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
- cik_sdma_copy_buffer(sctx, dst, src, dst_x, src_box->x, src_box->width);
- return;
- }
-
- /* Before re-enabling this, please make sure you can hit all newly
- * enabled paths in your testing, preferably with both piglit (in
- * particular the streaming-texture-leak test) and real world apps
- * (e.g. the UE4 Elemental demo).
- */
- goto fallback;
-
- if (src->format != dst->format ||
- rdst->surface.nsamples > 1 || rsrc->surface.nsamples > 1 ||
- (rdst->dirty_level_mask | rdst->stencil_dirty_level_mask) & (1 << dst_level) ||
- rdst->dcc_offset || rsrc->dcc_offset) {
- goto fallback;
- }
-
- if (rsrc->dirty_level_mask & (1 << src_level)) {
- if (rsrc->htile_buffer)
- goto fallback;
-
- ctx->flush_resource(ctx, src);
- }
-
- src_x = util_format_get_nblocksx(src->format, src_box->x);
- dst_x = util_format_get_nblocksx(src->format, dst_x);
- src_y = util_format_get_nblocksy(src->format, src_box->y);
- dst_y = util_format_get_nblocksy(src->format, dst_y);
-
- dst_pitch = rdst->surface.level[dst_level].pitch_bytes;
- src_pitch = rsrc->surface.level[src_level].pitch_bytes;
- src_w = rsrc->surface.level[src_level].npix_x;
- dst_w = rdst->surface.level[dst_level].npix_x;
-
- if (src_pitch != dst_pitch || src_box->x || dst_x || src_w != dst_w ||
- src_box->width != src_w ||
- rsrc->surface.level[src_level].nblk_y !=
- rdst->surface.level[dst_level].nblk_y) {
- /* FIXME CIK can do partial blit */
- goto fallback;
- }
-
- bpe = rdst->surface.bpe;
- copy_height = src_box->height / rsrc->surface.blk_h;
- dst_mode = rdst->surface.level[dst_level].mode;
- src_mode = rsrc->surface.level[src_level].mode;
- /* downcast linear aligned to linear to simplify test */
- src_mode = src_mode == RADEON_SURF_MODE_LINEAR_ALIGNED ? RADEON_SURF_MODE_LINEAR : src_mode;
- dst_mode = dst_mode == RADEON_SURF_MODE_LINEAR_ALIGNED ? RADEON_SURF_MODE_LINEAR : dst_mode;
-
- /* Dimensions must be aligned to (macro)tiles */
- switch (src_mode == RADEON_SURF_MODE_LINEAR ? dst_mode : src_mode) {
- case RADEON_SURF_MODE_1D:
- if ((src_x % 8) || (src_y % 8) || (dst_x % 8) || (dst_y % 8) ||
- (copy_height % 8))
- goto fallback;
- y_align = 8;
- break;
- case RADEON_SURF_MODE_2D: {
- unsigned mtilew, mtileh, num_banks;
-
- switch (si_num_banks(sctx->screen, rsrc)) {
- case V_02803C_ADDR_SURF_2_BANK:
- default:
- num_banks = 2;
- break;
- case V_02803C_ADDR_SURF_4_BANK:
- num_banks = 4;
- break;
- case V_02803C_ADDR_SURF_8_BANK:
- num_banks = 8;
- break;
- case V_02803C_ADDR_SURF_16_BANK:
- num_banks = 16;
- break;
- }
-
- mtilew = (8 * rsrc->surface.bankw *
- sctx->screen->b.info.num_tile_pipes) *
- rsrc->surface.mtilea;
- assert(!(mtilew & (mtilew - 1)));
- mtileh = (8 * rsrc->surface.bankh * num_banks) /
- rsrc->surface.mtilea;
- assert(!(mtileh & (mtileh - 1)));
-
- if ((src_x & (mtilew - 1)) || (src_y & (mtileh - 1)) ||
- (dst_x & (mtilew - 1)) || (dst_y & (mtileh - 1)) ||
- (copy_height & (mtileh - 1)))
- goto fallback;
-
- y_align = mtileh;
- break;
- }
- default:
- y_align = 1;
- }
-
- if (src_mode == dst_mode) {
- uint64_t dst_offset, src_offset;
- unsigned src_h, dst_h;
-
- src_h = rsrc->surface.level[src_level].npix_y;
- dst_h = rdst->surface.level[dst_level].npix_y;
-
- if (src_box->depth > 1 &&
- (src_y || dst_y || src_h != dst_h || src_box->height != src_h))
- goto fallback;
-
- /* simple dma blit would do NOTE code here assume :
- * dst_pitch == src_pitch
- */
- src_offset= rsrc->surface.level[src_level].offset;
- src_offset += rsrc->surface.level[src_level].slice_size * src_box->z;
- src_offset += src_y * src_pitch + src_x * bpe;
- dst_offset = rdst->surface.level[dst_level].offset;
- dst_offset += rdst->surface.level[dst_level].slice_size * dst_z;
- dst_offset += dst_y * dst_pitch + dst_x * bpe;
- cik_sdma_do_copy_buffer(sctx, dst, src, dst_offset, src_offset,
- src_box->depth *
- rsrc->surface.level[src_level].slice_size);
- } else {
- if (dst_y != src_y || src_box->depth > 1 || src_box->z || dst_z)
- goto fallback;
-
- cik_sdma_copy_tile(sctx, dst, dst_level, src, src_level,
- src_y, copy_height, y_align, dst_pitch, bpe);
- }
- return;
+ struct si_context *sctx = (struct si_context *)ctx;
+
+ assert(src->target != PIPE_BUFFER);
+
+ if (!sctx->sdma_cs || src->flags & PIPE_RESOURCE_FLAG_SPARSE ||
+ dst->flags & PIPE_RESOURCE_FLAG_SPARSE)
+ goto fallback;
+
+ /* SDMA causes corruption. See:
+ * https://bugs.freedesktop.org/show_bug.cgi?id=110575
+ * https://bugs.freedesktop.org/show_bug.cgi?id=110635
+ *
+ * Keep SDMA enabled on APUs.
+ */
+ if (sctx->screen->debug_flags & DBG(FORCE_SDMA) ||
+ (!sctx->screen->info.has_dedicated_vram &&
+ !(sctx->screen->debug_flags & DBG(NO_SDMA_COPY_IMAGE)))) {
+ if ((sctx->chip_class == GFX7 || sctx->chip_class == GFX8) &&
+ cik_sdma_copy_texture(sctx, dst, dst_level, dstx, dsty, dstz, src, src_level, src_box))
+ return;
+ else if (sctx->chip_class == GFX9 && si_sdma_v4_copy_texture(sctx, dst, dst_level, dstx, dsty,
+ dstz, src, src_level, src_box))
+ return;
+ }
fallback:
- si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
- src, src_level, src_box);
+ si_resource_copy_region(ctx, dst, dst_level, dstx, dsty, dstz, src, src_level, src_box);
+}
+
+void cik_init_sdma_functions(struct si_context *sctx)
+{
+ sctx->dma_copy = cik_sdma_copy;
}
projects / mesa.git / blobdiff