drm/radeon: overhaul texture checking. (v3)

the texture checking code didn't work for block formats like s3tc,
this overhauls it to work for all types.

v2: add texture array support.
v3: add subsampled formats

Signed-off-by: Dave Airlie <airlied@redhat.com>
This commit is contained in:
Dave Airlie 2011-02-18 05:51:58 +00:00
parent 82ef594efb
commit 60b212f8dd
2 changed files with 245 additions and 113 deletions

View File

@ -71,75 +71,164 @@ struct r600_cs_track {
u64 db_bo_mc;
};
#define FMT_8_BIT(fmt, vc) [fmt] = { 1, 1, 1, vc }
#define FMT_16_BIT(fmt, vc) [fmt] = { 1, 1, 2, vc }
#define FMT_24_BIT(fmt) [fmt] = { 1, 1, 3, 0 }
#define FMT_32_BIT(fmt, vc) [fmt] = { 1, 1, 4, vc }
#define FMT_48_BIT(fmt) [fmt] = { 1, 1, 6, 0 }
#define FMT_64_BIT(fmt, vc) [fmt] = { 1, 1, 8, vc }
#define FMT_96_BIT(fmt) [fmt] = { 1, 1, 12, 0 }
#define FMT_128_BIT(fmt, vc) [fmt] = { 1, 1, 16, vc }
struct gpu_formats {
unsigned blockwidth;
unsigned blockheight;
unsigned blocksize;
unsigned valid_color;
};
static const struct gpu_formats color_formats_table[] = {
/* 8 bit */
FMT_8_BIT(V_038004_COLOR_8, 1),
FMT_8_BIT(V_038004_COLOR_4_4, 1),
FMT_8_BIT(V_038004_COLOR_3_3_2, 1),
FMT_8_BIT(V_038004_FMT_1, 0),
/* 16-bit */
FMT_16_BIT(V_038004_COLOR_16, 1),
FMT_16_BIT(V_038004_COLOR_16_FLOAT, 1),
FMT_16_BIT(V_038004_COLOR_8_8, 1),
FMT_16_BIT(V_038004_COLOR_5_6_5, 1),
FMT_16_BIT(V_038004_COLOR_6_5_5, 1),
FMT_16_BIT(V_038004_COLOR_1_5_5_5, 1),
FMT_16_BIT(V_038004_COLOR_4_4_4_4, 1),
FMT_16_BIT(V_038004_COLOR_5_5_5_1, 1),
/* 24-bit */
FMT_24_BIT(V_038004_FMT_8_8_8),
/* 32-bit */
FMT_32_BIT(V_038004_COLOR_32, 1),
FMT_32_BIT(V_038004_COLOR_32_FLOAT, 1),
FMT_32_BIT(V_038004_COLOR_16_16, 1),
FMT_32_BIT(V_038004_COLOR_16_16_FLOAT, 1),
FMT_32_BIT(V_038004_COLOR_8_24, 1),
FMT_32_BIT(V_038004_COLOR_8_24_FLOAT, 1),
FMT_32_BIT(V_038004_COLOR_24_8, 1),
FMT_32_BIT(V_038004_COLOR_24_8_FLOAT, 1),
FMT_32_BIT(V_038004_COLOR_10_11_11, 1),
FMT_32_BIT(V_038004_COLOR_10_11_11_FLOAT, 1),
FMT_32_BIT(V_038004_COLOR_11_11_10, 1),
FMT_32_BIT(V_038004_COLOR_11_11_10_FLOAT, 1),
FMT_32_BIT(V_038004_COLOR_2_10_10_10, 1),
FMT_32_BIT(V_038004_COLOR_8_8_8_8, 1),
FMT_32_BIT(V_038004_COLOR_10_10_10_2, 1),
FMT_32_BIT(V_038004_FMT_5_9_9_9_SHAREDEXP, 0),
FMT_32_BIT(V_038004_FMT_32_AS_8, 0),
FMT_32_BIT(V_038004_FMT_32_AS_8_8, 0),
/* 48-bit */
FMT_48_BIT(V_038004_FMT_16_16_16),
FMT_48_BIT(V_038004_FMT_16_16_16_FLOAT),
/* 64-bit */
FMT_64_BIT(V_038004_COLOR_X24_8_32_FLOAT, 1),
FMT_64_BIT(V_038004_COLOR_32_32, 1),
FMT_64_BIT(V_038004_COLOR_32_32_FLOAT, 1),
FMT_64_BIT(V_038004_COLOR_16_16_16_16, 1),
FMT_64_BIT(V_038004_COLOR_16_16_16_16_FLOAT, 1),
FMT_96_BIT(V_038004_FMT_32_32_32),
FMT_96_BIT(V_038004_FMT_32_32_32_FLOAT),
/* 128-bit */
FMT_128_BIT(V_038004_COLOR_32_32_32_32, 1),
FMT_128_BIT(V_038004_COLOR_32_32_32_32_FLOAT, 1),
[V_038004_FMT_GB_GR] = { 2, 1, 4, 0 },
[V_038004_FMT_BG_RG] = { 2, 1, 4, 0 },
/* block compressed formats */
[V_038004_FMT_BC1] = { 4, 4, 8, 0 },
[V_038004_FMT_BC2] = { 4, 4, 16, 0 },
[V_038004_FMT_BC3] = { 4, 4, 16, 0 },
[V_038004_FMT_BC4] = { 4, 4, 8, 0 },
[V_038004_FMT_BC5] = { 4, 4, 16, 0},
};
static inline bool fmt_is_valid_color(u32 format)
{
if (format > ARRAY_SIZE(color_formats_table))
return false;
if (color_formats_table[format].valid_color)
return true;
return false;
}
static inline bool fmt_is_valid_texture(u32 format)
{
if (format > ARRAY_SIZE(color_formats_table))
return false;
if (color_formats_table[format].blockwidth > 0)
return true;
return false;
}
static inline int fmt_get_blocksize(u32 format)
{
if (format > ARRAY_SIZE(color_formats_table))
return 0;
return color_formats_table[format].blocksize;
}
static inline int fmt_get_nblocksx(u32 format, u32 w)
{
unsigned bw;
if (format > ARRAY_SIZE(color_formats_table))
return 0;
bw = color_formats_table[format].blockwidth;
if (bw == 0)
return 0;
return (w + bw - 1) / bw;
}
static inline int fmt_get_nblocksy(u32 format, u32 h)
{
unsigned bh;
if (format > ARRAY_SIZE(color_formats_table))
return 0;
bh = color_formats_table[format].blockheight;
if (bh == 0)
return 0;
return (h + bh - 1) / bh;
}
static inline int r600_bpe_from_format(u32 *bpe, u32 format)
{
switch (format) {
case V_038004_COLOR_8:
case V_038004_COLOR_4_4:
case V_038004_COLOR_3_3_2:
case V_038004_FMT_1:
*bpe = 1;
break;
case V_038004_COLOR_16:
case V_038004_COLOR_16_FLOAT:
case V_038004_COLOR_8_8:
case V_038004_COLOR_5_6_5:
case V_038004_COLOR_6_5_5:
case V_038004_COLOR_1_5_5_5:
case V_038004_COLOR_4_4_4_4:
case V_038004_COLOR_5_5_5_1:
*bpe = 2;
break;
case V_038004_FMT_8_8_8:
*bpe = 3;
break;
case V_038004_COLOR_32:
case V_038004_COLOR_32_FLOAT:
case V_038004_COLOR_16_16:
case V_038004_COLOR_16_16_FLOAT:
case V_038004_COLOR_8_24:
case V_038004_COLOR_8_24_FLOAT:
case V_038004_COLOR_24_8:
case V_038004_COLOR_24_8_FLOAT:
case V_038004_COLOR_10_11_11:
case V_038004_COLOR_10_11_11_FLOAT:
case V_038004_COLOR_11_11_10:
case V_038004_COLOR_11_11_10_FLOAT:
case V_038004_COLOR_2_10_10_10:
case V_038004_COLOR_8_8_8_8:
case V_038004_COLOR_10_10_10_2:
case V_038004_FMT_5_9_9_9_SHAREDEXP:
case V_038004_FMT_32_AS_8:
case V_038004_FMT_32_AS_8_8:
*bpe = 4;
break;
case V_038004_COLOR_X24_8_32_FLOAT:
case V_038004_COLOR_32_32:
case V_038004_COLOR_32_32_FLOAT:
case V_038004_COLOR_16_16_16_16:
case V_038004_COLOR_16_16_16_16_FLOAT:
*bpe = 8;
break;
case V_038004_FMT_16_16_16:
case V_038004_FMT_16_16_16_FLOAT:
*bpe = 6;
break;
case V_038004_FMT_32_32_32:
case V_038004_FMT_32_32_32_FLOAT:
*bpe = 12;
break;
case V_038004_COLOR_32_32_32_32:
case V_038004_COLOR_32_32_32_32_FLOAT:
*bpe = 16;
break;
case V_038004_FMT_GB_GR:
case V_038004_FMT_BG_RG:
case V_038004_COLOR_INVALID:
default:
*bpe = 16;
return -EINVAL;
}
unsigned res;
if (format > ARRAY_SIZE(color_formats_table))
goto fail;
res = color_formats_table[format].blocksize;
if (res == 0)
goto fail;
*bpe = res;
return 0;
fail:
*bpe = 16;
return -EINVAL;
}
struct array_mode_checker {
@ -148,7 +237,7 @@ struct array_mode_checker {
u32 nbanks;
u32 npipes;
u32 nsamples;
u32 bpe;
u32 blocksize;
};
/* returns alignment in pixels for pitch/height/depth and bytes for base */
@ -162,7 +251,7 @@ static inline int r600_get_array_mode_alignment(struct array_mode_checker *value
u32 tile_height = 8;
u32 macro_tile_width = values->nbanks;
u32 macro_tile_height = values->npipes;
u32 tile_bytes = tile_width * tile_height * values->bpe * values->nsamples;
u32 tile_bytes = tile_width * tile_height * values->blocksize * values->nsamples;
u32 macro_tile_bytes = macro_tile_width * macro_tile_height * tile_bytes;
switch (values->array_mode) {
@ -174,7 +263,7 @@ static inline int r600_get_array_mode_alignment(struct array_mode_checker *value
*base_align = 1;
break;
case ARRAY_LINEAR_ALIGNED:
*pitch_align = max((u32)64, (u32)(values->group_size / values->bpe));
*pitch_align = max((u32)64, (u32)(values->group_size / values->blocksize));
*height_align = tile_height;
*depth_align = 1;
*base_align = values->group_size;
@ -182,7 +271,7 @@ static inline int r600_get_array_mode_alignment(struct array_mode_checker *value
case ARRAY_1D_TILED_THIN1:
*pitch_align = max((u32)tile_width,
(u32)(values->group_size /
(tile_height * values->bpe * values->nsamples)));
(tile_height * values->blocksize * values->nsamples)));
*height_align = tile_height;
*depth_align = 1;
*base_align = values->group_size;
@ -190,12 +279,12 @@ static inline int r600_get_array_mode_alignment(struct array_mode_checker *value
case ARRAY_2D_TILED_THIN1:
*pitch_align = max((u32)macro_tile_width,
(u32)(((values->group_size / tile_height) /
(values->bpe * values->nsamples)) *
(values->blocksize * values->nsamples)) *
values->nbanks)) * tile_width;
*height_align = macro_tile_height * tile_height;
*depth_align = 1;
*base_align = max(macro_tile_bytes,
(*pitch_align) * values->bpe * (*height_align) * values->nsamples);
(*pitch_align) * values->blocksize * (*height_align) * values->nsamples);
break;
default:
return -EINVAL;
@ -234,21 +323,22 @@ static void r600_cs_track_init(struct r600_cs_track *track)
static inline int r600_cs_track_validate_cb(struct radeon_cs_parser *p, int i)
{
struct r600_cs_track *track = p->track;
u32 bpe = 0, slice_tile_max, size, tmp;
u32 slice_tile_max, size, tmp;
u32 height, height_align, pitch, pitch_align, depth_align;
u64 base_offset, base_align;
struct array_mode_checker array_check;
volatile u32 *ib = p->ib->ptr;
unsigned array_mode;
u32 format;
if (G_0280A0_TILE_MODE(track->cb_color_info[i])) {
dev_warn(p->dev, "FMASK or CMASK buffer are not supported by this kernel\n");
return -EINVAL;
}
size = radeon_bo_size(track->cb_color_bo[i]) - track->cb_color_bo_offset[i];
if (r600_bpe_from_format(&bpe, G_0280A0_FORMAT(track->cb_color_info[i]))) {
format = G_0280A0_FORMAT(track->cb_color_info[i]);
if (!fmt_is_valid_color(format)) {
dev_warn(p->dev, "%s:%d cb invalid format %d for %d (0x%08X)\n",
__func__, __LINE__, G_0280A0_FORMAT(track->cb_color_info[i]),
__func__, __LINE__, format,
i, track->cb_color_info[i]);
return -EINVAL;
}
@ -267,7 +357,7 @@ static inline int r600_cs_track_validate_cb(struct radeon_cs_parser *p, int i)
array_check.nbanks = track->nbanks;
array_check.npipes = track->npipes;
array_check.nsamples = track->nsamples;
array_check.bpe = bpe;
array_check.blocksize = fmt_get_blocksize(format);
if (r600_get_array_mode_alignment(&array_check,
&pitch_align, &height_align, &depth_align, &base_align)) {
dev_warn(p->dev, "%s invalid tiling %d for %d (0x%08X)\n", __func__,
@ -310,7 +400,7 @@ static inline int r600_cs_track_validate_cb(struct radeon_cs_parser *p, int i)
}
/* check offset */
tmp = height * pitch * bpe;
tmp = fmt_get_nblocksy(format, height) * fmt_get_nblocksx(format, pitch) * fmt_get_blocksize(format);
if ((tmp + track->cb_color_bo_offset[i]) > radeon_bo_size(track->cb_color_bo[i])) {
if (array_mode == V_0280A0_ARRAY_LINEAR_GENERAL) {
/* the initial DDX does bad things with the CB size occasionally */
@ -432,7 +522,7 @@ static int r600_cs_track_check(struct radeon_cs_parser *p)
array_check.nbanks = track->nbanks;
array_check.npipes = track->npipes;
array_check.nsamples = track->nsamples;
array_check.bpe = bpe;
array_check.blocksize = bpe;
if (r600_get_array_mode_alignment(&array_check,
&pitch_align, &height_align, &depth_align, &base_align)) {
dev_warn(p->dev, "%s invalid tiling %d (0x%08X)\n", __func__,
@ -1107,39 +1197,61 @@ static inline int r600_cs_check_reg(struct radeon_cs_parser *p, u32 reg, u32 idx
return 0;
}
static inline unsigned minify(unsigned size, unsigned levels)
static inline unsigned mip_minify(unsigned size, unsigned level)
{
size = size >> levels;
if (size < 1)
size = 1;
return size;
unsigned val;
val = max(1U, size >> level);
if (level > 0)
val = roundup_pow_of_two(val);
return val;
}
static void r600_texture_size(unsigned nfaces, unsigned blevel, unsigned nlevels,
unsigned w0, unsigned h0, unsigned d0, unsigned bpe,
unsigned pitch_align,
static void r600_texture_size(unsigned nfaces, unsigned blevel, unsigned llevel,
unsigned w0, unsigned h0, unsigned d0, unsigned format,
unsigned block_align, unsigned height_align, unsigned base_align,
unsigned *l0_size, unsigned *mipmap_size)
{
unsigned offset, i, level, face;
unsigned width, height, depth, rowstride, size;
unsigned offset, i, level;
unsigned width, height, depth, size;
unsigned blocksize;
unsigned nbx, nby;
unsigned nlevels = llevel - blevel + 1;
w0 = minify(w0, 0);
h0 = minify(h0, 0);
d0 = minify(d0, 0);
*l0_size = -1;
blocksize = fmt_get_blocksize(format);
w0 = mip_minify(w0, 0);
h0 = mip_minify(h0, 0);
d0 = mip_minify(d0, 0);
for(i = 0, offset = 0, level = blevel; i < nlevels; i++, level++) {
width = minify(w0, i);
height = minify(h0, i);
depth = minify(d0, i);
for(face = 0; face < nfaces; face++) {
rowstride = ALIGN((width * bpe), pitch_align);
size = height * rowstride * depth;
offset += size;
offset = (offset + 0x1f) & ~0x1f;
}
width = mip_minify(w0, i);
nbx = fmt_get_nblocksx(format, width);
nbx = round_up(nbx, block_align);
height = mip_minify(h0, i);
nby = fmt_get_nblocksy(format, height);
nby = round_up(nby, height_align);
depth = mip_minify(d0, i);
size = nbx * nby * blocksize;
if (nfaces)
size *= nfaces;
else
size *= depth;
if (i == 0)
*l0_size = size;
if (i == 0 || i == 1)
offset = round_up(offset, base_align);
offset += size;
}
*l0_size = ALIGN((w0 * bpe), pitch_align) * h0 * d0;
*mipmap_size = offset;
if (!nlevels)
if (llevel == 0)
*mipmap_size = *l0_size;
if (!blevel)
*mipmap_size -= *l0_size;
@ -1163,11 +1275,13 @@ static inline int r600_check_texture_resource(struct radeon_cs_parser *p, u32 i
u32 tiling_flags)
{
struct r600_cs_track *track = p->track;
u32 nfaces, nlevels, blevel, w0, h0, d0, bpe = 0;
u32 nfaces, llevel, blevel, w0, h0, d0;
u32 word0, word1, l0_size, mipmap_size;
u32 height_align, pitch, pitch_align, depth_align;
u32 array, barray, larray;
u64 base_align;
struct array_mode_checker array_check;
u32 format;
/* on legacy kernel we don't perform advanced check */
if (p->rdev == NULL)
@ -1193,19 +1307,25 @@ static inline int r600_check_texture_resource(struct radeon_cs_parser *p, u32 i
case V_038000_SQ_TEX_DIM_3D:
break;
case V_038000_SQ_TEX_DIM_CUBEMAP:
nfaces = 6;
if (p->family >= CHIP_RV770)
nfaces = 8;
else
nfaces = 6;
break;
case V_038000_SQ_TEX_DIM_1D_ARRAY:
case V_038000_SQ_TEX_DIM_2D_ARRAY:
array = 1;
break;
case V_038000_SQ_TEX_DIM_2D_MSAA:
case V_038000_SQ_TEX_DIM_2D_ARRAY_MSAA:
default:
dev_warn(p->dev, "this kernel doesn't support %d texture dim\n", G_038000_DIM(word0));
return -EINVAL;
}
if (r600_bpe_from_format(&bpe, G_038004_DATA_FORMAT(word1))) {
format = G_038004_DATA_FORMAT(word1);
if (!fmt_is_valid_texture(format)) {
dev_warn(p->dev, "%s:%d texture invalid format %d\n",
__func__, __LINE__, G_038004_DATA_FORMAT(word1));
__func__, __LINE__, format);
return -EINVAL;
}
@ -1216,7 +1336,7 @@ static inline int r600_check_texture_resource(struct radeon_cs_parser *p, u32 i
array_check.nbanks = track->nbanks;
array_check.npipes = track->npipes;
array_check.nsamples = 1;
array_check.bpe = bpe;
array_check.blocksize = fmt_get_blocksize(format);
if (r600_get_array_mode_alignment(&array_check,
&pitch_align, &height_align, &depth_align, &base_align)) {
dev_warn(p->dev, "%s:%d tex array mode (%d) invalid\n",
@ -1245,22 +1365,29 @@ static inline int r600_check_texture_resource(struct radeon_cs_parser *p, u32 i
word0 = radeon_get_ib_value(p, idx + 4);
word1 = radeon_get_ib_value(p, idx + 5);
blevel = G_038010_BASE_LEVEL(word0);
nlevels = G_038014_LAST_LEVEL(word1);
r600_texture_size(nfaces, blevel, nlevels, w0, h0, d0, bpe,
(pitch_align * bpe),
llevel = G_038014_LAST_LEVEL(word1);
if (array == 1) {
barray = G_038014_BASE_ARRAY(word1);
larray = G_038014_LAST_ARRAY(word1);
nfaces = larray - barray + 1;
}
r600_texture_size(nfaces, blevel, llevel, w0, h0, d0, format,
pitch_align, height_align, base_align,
&l0_size, &mipmap_size);
/* using get ib will give us the offset into the texture bo */
word0 = radeon_get_ib_value(p, idx + 2) << 8;
if ((l0_size + word0) > radeon_bo_size(texture)) {
dev_warn(p->dev, "texture bo too small (%d %d %d %d -> %d have %ld)\n",
w0, h0, bpe, word0, l0_size, radeon_bo_size(texture));
w0, h0, format, word0, l0_size, radeon_bo_size(texture));
dev_warn(p->dev, "alignments %d %d %d %lld\n", pitch, pitch_align, height_align, base_align);
return -EINVAL;
}
/* using get ib will give us the offset into the mipmap bo */
word0 = radeon_get_ib_value(p, idx + 3) << 8;
if ((mipmap_size + word0) > radeon_bo_size(mipmap)) {
/*dev_warn(p->dev, "mipmap bo too small (%d %d %d %d %d %d -> %d have %ld)\n",
w0, h0, bpe, blevel, nlevels, word0, mipmap_size, radeon_bo_size(texture));*/
w0, h0, format, blevel, nlevels, word0, mipmap_size, radeon_bo_size(texture));*/
}
return 0;
}

View File

@ -1303,6 +1303,11 @@
#define V_038004_FMT_16_16_16_FLOAT 0x0000002E
#define V_038004_FMT_32_32_32 0x0000002F
#define V_038004_FMT_32_32_32_FLOAT 0x00000030
#define V_038004_FMT_BC1 0x00000031
#define V_038004_FMT_BC2 0x00000032
#define V_038004_FMT_BC3 0x00000033
#define V_038004_FMT_BC4 0x00000034
#define V_038004_FMT_BC5 0x00000035
#define R_038010_SQ_TEX_RESOURCE_WORD4_0 0x038010
#define S_038010_FORMAT_COMP_X(x) (((x) & 0x3) << 0)
#define G_038010_FORMAT_COMP_X(x) (((x) >> 0) & 0x3)