[Mesa-dev] [PATCH 3/6] radeonsi: don't change pipe_resource in resource_copy_region
Marek Olšák
maraeo at gmail.com
Thu Jul 16 09:55:28 PDT 2015
From: Marek Olšák <marek.olsak at amd.com>
Copied from r600g. pipe_resource can be shared by multiple threads, so we
shouldn't change it.
---
src/gallium/drivers/radeon/r600_pipe_common.h | 1 -
src/gallium/drivers/radeonsi/si_blit.c | 201 +++++++-------------------
src/gallium/drivers/radeonsi/si_state.c | 62 ++++++--
src/gallium/drivers/radeonsi/si_state.h | 6 +
4 files changed, 110 insertions(+), 160 deletions(-)
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index a471426..aeb1e7a 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -214,7 +214,6 @@ struct r600_texture {
float depth_clear_value;
bool non_disp_tiling; /* R600-Cayman only */
- unsigned mipmap_shift;
};
struct r600_surface {
diff --git a/src/gallium/drivers/radeonsi/si_blit.c b/src/gallium/drivers/radeonsi/si_blit.c
index 6c7b383..f6db3f5 100644
--- a/src/gallium/drivers/radeonsi/si_blit.c
+++ b/src/gallium/drivers/radeonsi/si_blit.c
@@ -455,89 +455,6 @@ struct texture_orig_info {
unsigned npix0_y;
};
-static void si_compressed_to_blittable(struct pipe_resource *tex,
- unsigned level,
- struct texture_orig_info *orig)
-{
- struct r600_texture *rtex = (struct r600_texture*)tex;
- unsigned pixsize = util_format_get_blocksize(rtex->resource.b.b.format);
- int new_format;
- int new_height, new_width;
-
- orig->format = tex->format;
- orig->width0 = tex->width0;
- orig->height0 = tex->height0;
- orig->npix0_x = rtex->surface.level[0].npix_x;
- orig->npix0_y = rtex->surface.level[0].npix_y;
- orig->npix_x = rtex->surface.level[level].npix_x;
- orig->npix_y = rtex->surface.level[level].npix_y;
-
- if (pixsize == 8)
- new_format = PIPE_FORMAT_R16G16B16A16_UINT; /* 64-bit block */
- else
- new_format = PIPE_FORMAT_R32G32B32A32_UINT; /* 128-bit block */
-
- new_width = util_format_get_nblocksx(tex->format, orig->width0);
- new_height = util_format_get_nblocksy(tex->format, orig->height0);
-
- tex->width0 = new_width;
- tex->height0 = new_height;
- tex->format = new_format;
- rtex->surface.level[0].npix_x = util_format_get_nblocksx(orig->format, orig->npix0_x);
- rtex->surface.level[0].npix_y = util_format_get_nblocksy(orig->format, orig->npix0_y);
- rtex->surface.level[level].npix_x = util_format_get_nblocksx(orig->format, orig->npix_x);
- rtex->surface.level[level].npix_y = util_format_get_nblocksy(orig->format, orig->npix_y);
-
- /* By dividing the dimensions by 4, we effectively decrement
- * last_level by 2, therefore the last 2 mipmap levels disappear and
- * aren't blittable. Note that the last 3 mipmap levels (4x4, 2x2,
- * 1x1) have equal slice sizes, which is an important assumption
- * for this to work.
- *
- * In order to make the last 2 mipmap levels blittable, we have to
- * add the slice size of the last mipmap level to the texture
- * address, so that even though the hw thinks it reads last_level-2,
- * it will actually read last_level-1, and if we add the slice size*2,
- * it will read last_level. That's how this workaround works.
- */
- if (level > rtex->resource.b.b.last_level-2)
- rtex->mipmap_shift = level - (rtex->resource.b.b.last_level-2);
-}
-
-static void si_change_format(struct pipe_resource *tex,
- unsigned level,
- struct texture_orig_info *orig,
- enum pipe_format format)
-{
- struct r600_texture *rtex = (struct r600_texture*)tex;
-
- orig->format = tex->format;
- orig->width0 = tex->width0;
- orig->height0 = tex->height0;
- orig->npix0_x = rtex->surface.level[0].npix_x;
- orig->npix0_y = rtex->surface.level[0].npix_y;
- orig->npix_x = rtex->surface.level[level].npix_x;
- orig->npix_y = rtex->surface.level[level].npix_y;
-
- tex->format = format;
-}
-
-static void si_reset_blittable_to_orig(struct pipe_resource *tex,
- unsigned level,
- struct texture_orig_info *orig)
-{
- struct r600_texture *rtex = (struct r600_texture*)tex;
-
- tex->format = orig->format;
- tex->width0 = orig->width0;
- tex->height0 = orig->height0;
- rtex->surface.level[0].npix_x = orig->npix0_x;
- rtex->surface.level[0].npix_y = orig->npix0_y;
- rtex->surface.level[level].npix_x = orig->npix_x;
- rtex->surface.level[level].npix_y = orig->npix_y;
- rtex->mipmap_shift = 0;
-}
-
void si_resource_copy_region(struct pipe_context *ctx,
struct pipe_resource *dst,
unsigned dst_level,
@@ -547,114 +464,114 @@ void si_resource_copy_region(struct pipe_context *ctx,
const struct pipe_box *src_box)
{
struct si_context *sctx = (struct si_context *)ctx;
- struct r600_texture *rdst = (struct r600_texture*)dst;
struct pipe_surface *dst_view, dst_templ;
struct pipe_sampler_view src_templ, *src_view;
- struct texture_orig_info orig_info[2];
+ unsigned dst_width, dst_height, src_width0, src_height0;
+ unsigned src_force_level = 0;
struct pipe_box sbox, dstbox;
- boolean restore_orig[2];
- /* Fallback for buffers. */
+ /* Handle buffers first. */
if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
si_copy_buffer(sctx, dst, src, dstx, src_box->x, src_box->width, false);
return;
}
- memset(orig_info, 0, sizeof(orig_info));
+ assert(u_max_sample(dst) == u_max_sample(src));
/* The driver doesn't decompress resources automatically while
* u_blitter is rendering. */
si_decompress_subresource(ctx, src, src_level,
src_box->z, src_box->z + src_box->depth - 1);
- restore_orig[0] = restore_orig[1] = FALSE;
+ dst_width = u_minify(dst->width0, dst_level);
+ dst_height = u_minify(dst->height0, dst_level);
+ src_width0 = src->width0;
+ src_height0 = src->height0;
+
+ util_blitter_default_dst_texture(&dst_templ, dst, dst_level, dstz);
+ util_blitter_default_src_texture(&src_templ, src, src_level);
if (util_format_is_compressed(src->format) &&
util_format_is_compressed(dst->format)) {
- si_compressed_to_blittable(src, src_level, &orig_info[0]);
- restore_orig[0] = TRUE;
- sbox.x = util_format_get_nblocksx(orig_info[0].format, src_box->x);
- sbox.y = util_format_get_nblocksy(orig_info[0].format, src_box->y);
+ unsigned blocksize = util_format_get_blocksize(src->format);
+
+ if (blocksize == 8)
+ src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; /* 64-bit block */
+ else
+ src_templ.format = PIPE_FORMAT_R32G32B32A32_UINT; /* 128-bit block */
+ dst_templ.format = src_templ.format;
+
+ dst_width = util_format_get_nblocksx(dst->format, dst_width);
+ dst_height = util_format_get_nblocksy(dst->format, dst_height);
+ src_width0 = util_format_get_nblocksx(src->format, src_width0);
+ src_height0 = util_format_get_nblocksy(src->format, src_height0);
+
+ dstx = util_format_get_nblocksx(dst->format, dstx);
+ dsty = util_format_get_nblocksy(dst->format, dsty);
+
+ sbox.x = util_format_get_nblocksx(src->format, src_box->x);
+ sbox.y = util_format_get_nblocksy(src->format, src_box->y);
sbox.z = src_box->z;
- sbox.width = util_format_get_nblocksx(orig_info[0].format, src_box->width);
- sbox.height = util_format_get_nblocksy(orig_info[0].format, src_box->height);
+ sbox.width = util_format_get_nblocksx(src->format, src_box->width);
+ sbox.height = util_format_get_nblocksy(src->format, src_box->height);
sbox.depth = src_box->depth;
src_box = &sbox;
- si_compressed_to_blittable(dst, dst_level, &orig_info[1]);
- restore_orig[1] = TRUE;
- /* translate the dst box as well */
- dstx = util_format_get_nblocksx(orig_info[1].format, dstx);
- dsty = util_format_get_nblocksy(orig_info[1].format, dsty);
+ src_force_level = src_level;
} else if (!util_blitter_is_copy_supported(sctx->blitter, dst, src)) {
if (util_format_is_subsampled_422(src->format)) {
- /* XXX untested */
- si_change_format(src, src_level, &orig_info[0],
- PIPE_FORMAT_R8G8B8A8_UINT);
- si_change_format(dst, dst_level, &orig_info[1],
- PIPE_FORMAT_R8G8B8A8_UINT);
+ src_templ.format = PIPE_FORMAT_R8G8B8A8_UINT;
+ dst_templ.format = PIPE_FORMAT_R8G8B8A8_UINT;
+
+ dst_width = util_format_get_nblocksx(dst->format, dst_width);
+ src_width0 = util_format_get_nblocksx(src->format, src_width0);
+
+ dstx = util_format_get_nblocksx(dst->format, dstx);
sbox = *src_box;
- sbox.x = util_format_get_nblocksx(orig_info[0].format, src_box->x);
- sbox.width = util_format_get_nblocksx(orig_info[0].format, src_box->width);
+ sbox.x = util_format_get_nblocksx(src->format, src_box->x);
+ sbox.width = util_format_get_nblocksx(src->format, src_box->width);
src_box = &sbox;
- dstx = util_format_get_nblocksx(orig_info[1].format, dstx);
-
- restore_orig[0] = TRUE;
- restore_orig[1] = TRUE;
} else {
unsigned blocksize = util_format_get_blocksize(src->format);
switch (blocksize) {
case 1:
- si_change_format(src, src_level, &orig_info[0],
- PIPE_FORMAT_R8_UNORM);
- si_change_format(dst, dst_level, &orig_info[1],
- PIPE_FORMAT_R8_UNORM);
+ dst_templ.format = PIPE_FORMAT_R8_UNORM;
+ src_templ.format = PIPE_FORMAT_R8_UNORM;
break;
case 2:
- si_change_format(src, src_level, &orig_info[0],
- PIPE_FORMAT_R8G8_UNORM);
- si_change_format(dst, dst_level, &orig_info[1],
- PIPE_FORMAT_R8G8_UNORM);
+ dst_templ.format = PIPE_FORMAT_R8G8_UNORM;
+ src_templ.format = PIPE_FORMAT_R8G8_UNORM;
break;
case 4:
- si_change_format(src, src_level, &orig_info[0],
- PIPE_FORMAT_R8G8B8A8_UNORM);
- si_change_format(dst, dst_level, &orig_info[1],
- PIPE_FORMAT_R8G8B8A8_UNORM);
+ dst_templ.format = PIPE_FORMAT_R8G8B8A8_UNORM;
+ src_templ.format = PIPE_FORMAT_R8G8B8A8_UNORM;
break;
case 8:
- si_change_format(src, src_level, &orig_info[0],
- PIPE_FORMAT_R16G16B16A16_UINT);
- si_change_format(dst, dst_level, &orig_info[1],
- PIPE_FORMAT_R16G16B16A16_UINT);
+ dst_templ.format = PIPE_FORMAT_R16G16B16A16_UINT;
+ src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT;
break;
case 16:
- si_change_format(src, src_level, &orig_info[0],
- PIPE_FORMAT_R32G32B32A32_UINT);
- si_change_format(dst, dst_level, &orig_info[1],
- PIPE_FORMAT_R32G32B32A32_UINT);
+ dst_templ.format = PIPE_FORMAT_R32G32B32A32_UINT;
+ src_templ.format = PIPE_FORMAT_R32G32B32A32_UINT;
break;
default:
fprintf(stderr, "Unhandled format %s with blocksize %u\n",
util_format_short_name(src->format), blocksize);
assert(0);
}
- restore_orig[0] = TRUE;
- restore_orig[1] = TRUE;
}
}
/* Initialize the surface. */
- util_blitter_default_dst_texture(&dst_templ, dst, dst_level, dstz);
dst_view = r600_create_surface_custom(ctx, dst, &dst_templ,
- rdst->surface.level[dst_level].npix_x,
- rdst->surface.level[dst_level].npix_y);
+ dst_width, dst_height);
/* Initialize the sampler view. */
- util_blitter_default_src_texture(&src_templ, src, src_level);
- src_view = ctx->create_sampler_view(ctx, src, &src_templ);
+ src_view = si_create_sampler_view_custom(ctx, src, &src_templ,
+ src_width0, src_height0,
+ src_force_level);
u_box_3d(dstx, dsty, dstz, abs(src_box->width), abs(src_box->height),
abs(src_box->depth), &dstbox);
@@ -662,18 +579,12 @@ void si_resource_copy_region(struct pipe_context *ctx,
/* Copy. */
si_blitter_begin(ctx, SI_COPY);
util_blitter_blit_generic(sctx->blitter, dst_view, &dstbox,
- src_view, src_box, src->width0, src->height0,
+ src_view, src_box, src_width0, src_height0,
PIPE_MASK_RGBAZS, PIPE_TEX_FILTER_NEAREST, NULL);
si_blitter_end(ctx);
pipe_surface_reference(&dst_view, NULL);
pipe_sampler_view_reference(&src_view, NULL);
-
- if (restore_orig[0])
- si_reset_blittable_to_orig(src, src_level, &orig_info[0]);
-
- if (restore_orig[1])
- si_reset_blittable_to_orig(dst, dst_level, &orig_info[1]);
}
/* For MSAA integer resolving to work, we change the format to NORM using this function. */
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 0dd08a2..1d4a4e8 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -2272,15 +2272,28 @@ static void si_set_min_samples(struct pipe_context *ctx, unsigned min_samples)
* Samplers
*/
-static struct pipe_sampler_view *si_create_sampler_view(struct pipe_context *ctx,
- struct pipe_resource *texture,
- const struct pipe_sampler_view *state)
+/**
+ * Create a sampler view.
+ *
+ * @param ctx context
+ * @param texture texture
+ * @param state sampler view template
+ * @param width0 width0 override (for compressed textures as int)
+ * @param height0 height0 override (for compressed textures as int)
+ * @param force_level set the base address to the level (for compressed textures)
+ */
+struct pipe_sampler_view *
+si_create_sampler_view_custom(struct pipe_context *ctx,
+ struct pipe_resource *texture,
+ const struct pipe_sampler_view *state,
+ unsigned width0, unsigned height0,
+ unsigned force_level)
{
struct si_context *sctx = (struct si_context*)ctx;
struct si_sampler_view *view = CALLOC_STRUCT(si_sampler_view);
struct r600_texture *tmp = (struct r600_texture*)texture;
const struct util_format_description *desc;
- unsigned format, num_format;
+ unsigned format, num_format, base_level, first_level, last_level;
uint32_t pitch = 0;
unsigned char state_swizzle[4], swizzle[4];
unsigned height, depth, width;
@@ -2453,13 +2466,25 @@ static struct pipe_sampler_view *si_create_sampler_view(struct pipe_context *ctx
format = 0;
}
- /* not supported any more */
- //endian = si_colorformat_endian_swap(format);
+ base_level = 0;
+ first_level = state->u.tex.first_level;
+ last_level = state->u.tex.last_level;
+ width = width0;
+ height = height0;
+ depth = texture->depth0;
+
+ if (force_level) {
+ assert(force_level == first_level &&
+ force_level == last_level);
+ base_level = force_level;
+ first_level = 0;
+ last_level = 0;
+ width = u_minify(width, force_level);
+ height = u_minify(height, force_level);
+ depth = u_minify(depth, force_level);
+ }
- width = surflevel[0].npix_x;
- height = surflevel[0].npix_y;
- depth = surflevel[0].npix_z;
- pitch = surflevel[0].nblk_x * util_format_get_blockwidth(pipe_format);
+ pitch = surflevel[base_level].nblk_x * util_format_get_blockwidth(pipe_format);
if (texture->target == PIPE_TEXTURE_1D_ARRAY) {
height = 1;
@@ -2469,8 +2494,7 @@ static struct pipe_sampler_view *si_create_sampler_view(struct pipe_context *ctx
} else if (texture->target == PIPE_TEXTURE_CUBE_ARRAY)
depth = texture->array_size / 6;
- va = tmp->resource.gpu_address + surflevel[0].offset;
- va += tmp->mipmap_shift * surflevel[texture->last_level].slice_size * tmp->surface.array_size;
+ va = tmp->resource.gpu_address + surflevel[base_level].offset;
view->state[0] = va >> 8;
view->state[1] = (S_008F14_BASE_ADDRESS_HI(va >> 40) |
@@ -2483,10 +2507,10 @@ static struct pipe_sampler_view *si_create_sampler_view(struct pipe_context *ctx
S_008F1C_DST_SEL_Z(si_map_swizzle(swizzle[2])) |
S_008F1C_DST_SEL_W(si_map_swizzle(swizzle[3])) |
S_008F1C_BASE_LEVEL(texture->nr_samples > 1 ?
- 0 : state->u.tex.first_level - tmp->mipmap_shift) |
+ 0 : first_level) |
S_008F1C_LAST_LEVEL(texture->nr_samples > 1 ?
util_logbase2(texture->nr_samples) :
- state->u.tex.last_level - tmp->mipmap_shift) |
+ last_level) |
S_008F1C_TILING_INDEX(si_tile_mode_index(tmp, 0, false)) |
S_008F1C_POW2_PAD(texture->last_level > 0) |
S_008F1C_TYPE(si_tex_dim(texture->target, texture->nr_samples)));
@@ -2539,6 +2563,16 @@ static struct pipe_sampler_view *si_create_sampler_view(struct pipe_context *ctx
return &view->base;
}
+static struct pipe_sampler_view *
+si_create_sampler_view(struct pipe_context *ctx,
+ struct pipe_resource *texture,
+ const struct pipe_sampler_view *state)
+{
+ return si_create_sampler_view_custom(ctx, texture, state,
+ texture ? texture->width0 : 0,
+ texture ? texture->height0 : 0, 0);
+}
+
static void si_sampler_view_destroy(struct pipe_context *ctx,
struct pipe_sampler_view *state)
{
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index d1f2dff..634f947 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -264,6 +264,12 @@ unsigned cik_tile_split(unsigned tile_split);
unsigned si_array_mode(unsigned mode);
uint32_t si_num_banks(struct si_screen *sscreen, struct r600_texture *tex);
unsigned si_tile_mode_index(struct r600_texture *rtex, unsigned level, bool stencil);
+struct pipe_sampler_view *
+si_create_sampler_view_custom(struct pipe_context *ctx,
+ struct pipe_resource *texture,
+ const struct pipe_sampler_view *state,
+ unsigned width0, unsigned height0,
+ unsigned force_level);
/* si_state_shader.c */
void si_update_shaders(struct si_context *sctx);
--
2.1.0
More information about the mesa-dev
mailing list