#include "radeon/r600_cs.h"
#include "si_pipe.h"
#include "sid.h"
+#include "gfx9d.h"
#include "util/u_format.h"
#include "util/u_memory.h"
desc->shader_userdata_offset = shader_userdata_index * 4;
if (ce_offset) {
+ desc->uses_ce = true;
desc->ce_offset = *ce_offset;
/* make sure that ce_offset stays 32 byte aligned */
unsigned *out_offset, struct r600_resource **out_buf) {
uint64_t va;
- u_suballocator_alloc(sctx->ce_suballocator, size, 64, out_offset,
- (struct pipe_resource**)out_buf);
+ u_suballocator_alloc(sctx->ce_suballocator, size,
+ sctx->screen->b.info.tcc_cache_line_size,
+ out_offset, (struct pipe_resource**)out_buf);
if (!out_buf)
return false;
if (!desc->dirty_mask)
return true;
- if (sctx->ce_ib) {
+ if (sctx->ce_ib && desc->uses_ce) {
uint32_t const* list = (uint32_t const*)desc->list;
if (desc->ce_ram_dirty)
} else {
void *ptr;
- u_upload_alloc(sctx->b.uploader, 0, list_size, 256,
- &desc->buffer_offset,
- (struct pipe_resource**)&desc->buffer, &ptr);
+ u_upload_alloc(sctx->b.b.const_uploader, 0, list_size,
+ sctx->screen->b.info.tcc_cache_line_size,
+ &desc->buffer_offset,
+ (struct pipe_resource**)&desc->buffer, &ptr);
if (!desc->buffer)
return false; /* skip the draw call */
util_memcpy_cpu_to_le32(ptr, desc->list, list_size);
+ desc->gpu_list = ptr;
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, desc->buffer,
RADEON_USAGE_READ, RADEON_PRIO_DESCRIPTORS);
}
- desc->pointer_dirty = true;
desc->dirty_mask = 0;
if (atom)
if (resource->target == PIPE_BUFFER)
return;
- /* Now add separate DCC if it's present. */
+ /* Now add separate DCC or HTILE. */
rtex = (struct r600_texture*)resource;
- if (!rtex->dcc_separate_buffer)
- return;
+ if (rtex->dcc_separate_buffer) {
+ radeon_add_to_buffer_list_check_mem(&sctx->b, &sctx->b.gfx,
+ rtex->dcc_separate_buffer, usage,
+ RADEON_PRIO_DCC, check_mem);
+ }
- radeon_add_to_buffer_list_check_mem(&sctx->b, &sctx->b.gfx,
- rtex->dcc_separate_buffer, usage,
- RADEON_PRIO_DCC, check_mem);
+ if (rtex->htile_buffer &&
+ rtex->tc_compatible_htile &&
+ !is_stencil_sampler) {
+ radeon_add_to_buffer_list_check_mem(&sctx->b, &sctx->b.gfx,
+ rtex->htile_buffer, usage,
+ RADEON_PRIO_HTILE, check_mem);
+ }
}
static void si_sampler_views_begin_new_cs(struct si_context *sctx,
* \param is_stencil select between separate Z & Stencil
* \param state descriptor to update
*/
-void si_set_mutable_tex_desc_fields(struct r600_texture *tex,
- const struct radeon_surf_level *base_level_info,
+void si_set_mutable_tex_desc_fields(struct si_screen *sscreen,
+ struct r600_texture *tex,
+ const struct legacy_surf_level *base_level_info,
unsigned base_level, unsigned first_level,
unsigned block_width, bool is_stencil,
uint32_t *state)
{
- uint64_t va;
- unsigned pitch = base_level_info->nblk_x * block_width;
+ uint64_t va, meta_va = 0;
if (tex->is_depth && !r600_can_sample_zs(tex, is_stencil)) {
tex = tex->flushed_depth_texture;
is_stencil = false;
}
- va = tex->resource.gpu_address + base_level_info->offset;
+ va = tex->resource.gpu_address;
- state[1] &= C_008F14_BASE_ADDRESS_HI;
- state[3] &= C_008F1C_TILING_INDEX;
- state[4] &= C_008F20_PITCH;
- state[6] &= C_008F28_COMPRESSION_EN;
+ if (sscreen->b.chip_class >= GFX9) {
+ /* Only stencil_offset needs to be added here. */
+ if (is_stencil)
+ va += tex->surface.u.gfx9.stencil_offset;
+ else
+ va += tex->surface.u.gfx9.surf_offset;
+ } else {
+ va += base_level_info->offset;
+ }
state[0] = va >> 8;
+ state[1] &= C_008F14_BASE_ADDRESS_HI;
state[1] |= S_008F14_BASE_ADDRESS_HI(va >> 40);
- state[3] |= S_008F1C_TILING_INDEX(si_tile_mode_index(tex, base_level,
- is_stencil));
- state[4] |= S_008F20_PITCH(pitch - 1);
-
- if (tex->dcc_offset && first_level < tex->surface.num_dcc_levels) {
- state[6] |= S_008F28_COMPRESSION_EN(1);
- state[7] = ((!tex->dcc_separate_buffer ? tex->resource.gpu_address : 0) +
- tex->dcc_offset +
- base_level_info->dcc_offset) >> 8;
- } else if (tex->tc_compatible_htile) {
- state[6] |= S_008F28_COMPRESSION_EN(1);
- state[7] = tex->htile_buffer->gpu_address >> 8;
+
+ if (sscreen->b.chip_class >= VI) {
+ state[6] &= C_008F28_COMPRESSION_EN;
+ state[7] = 0;
+
+ if (vi_dcc_enabled(tex, first_level)) {
+ meta_va = (!tex->dcc_separate_buffer ? tex->resource.gpu_address : 0) +
+ tex->dcc_offset;
+
+ if (sscreen->b.chip_class <= VI)
+ meta_va += base_level_info->dcc_offset;
+ } else if (tex->tc_compatible_htile && !is_stencil) {
+ meta_va = tex->htile_buffer->gpu_address;
+ }
+
+ if (meta_va) {
+ state[6] |= S_008F28_COMPRESSION_EN(1);
+ state[7] = meta_va >> 8;
+ }
+ }
+
+ if (sscreen->b.chip_class >= GFX9) {
+ state[3] &= C_008F1C_SW_MODE;
+ state[4] &= C_008F20_PITCH_GFX9;
+
+ if (is_stencil) {
+ state[3] |= S_008F1C_SW_MODE(tex->surface.u.gfx9.stencil.swizzle_mode);
+ state[4] |= S_008F20_PITCH_GFX9(tex->surface.u.gfx9.stencil.epitch);
+ } else {
+ state[3] |= S_008F1C_SW_MODE(tex->surface.u.gfx9.surf.swizzle_mode);
+ state[4] |= S_008F20_PITCH_GFX9(tex->surface.u.gfx9.surf.epitch);
+ }
+
+ state[5] &= C_008F24_META_DATA_ADDRESS &
+ C_008F24_META_PIPE_ALIGNED &
+ C_008F24_META_RB_ALIGNED;
+ if (meta_va) {
+ struct gfx9_surf_meta_flags meta;
+
+ if (tex->dcc_offset)
+ meta = tex->surface.u.gfx9.dcc;
+ else
+ meta = tex->surface.u.gfx9.htile;
+
+ state[5] |= S_008F24_META_DATA_ADDRESS(meta_va >> 40) |
+ S_008F24_META_PIPE_ALIGNED(meta.pipe_aligned) |
+ S_008F24_META_RB_ALIGNED(meta.rb_aligned);
+ }
+ } else {
+ /* SI-CI-VI */
+ unsigned pitch = base_level_info->nblk_x * block_width;
+ unsigned index = si_tile_mode_index(tex, base_level, is_stencil);
+
+ state[3] &= C_008F1C_TILING_INDEX;
+ state[3] |= S_008F1C_TILING_INDEX(index);
+ state[4] &= C_008F20_PITCH_GFX6;
+ state[4] |= S_008F20_PITCH_GFX6(pitch - 1);
}
}
struct si_sampler_views *views = &sctx->samplers[shader].views;
struct si_sampler_view *rview = (struct si_sampler_view*)view;
struct si_descriptors *descs = si_sampler_descriptors(sctx, shader);
+ uint32_t *desc = descs->list + slot * 16;
if (views->views[slot] == view && !disallow_early_out)
return;
if (view) {
struct r600_texture *rtex = (struct r600_texture *)view->texture;
- uint32_t *desc = descs->list + slot * 16;
+ bool is_buffer = rtex->resource.b.b.target == PIPE_BUFFER;
+
+ if (unlikely(!is_buffer && rview->dcc_incompatible)) {
+ if (vi_dcc_enabled(rtex, view->u.tex.first_level))
+ if (!r600_texture_disable_dcc(&sctx->b, rtex))
+ sctx->b.decompress_dcc(&sctx->b.b, rtex);
+
+ rview->dcc_incompatible = false;
+ }
assert(rtex); /* views with texture == NULL aren't supported */
pipe_sampler_view_reference(&views->views[slot], view);
memcpy(desc, rview->state, 8*4);
- if (rtex->resource.b.b.target == PIPE_BUFFER) {
+ if (is_buffer) {
rtex->resource.bind_history |= PIPE_BIND_SAMPLER_VIEW;
si_set_buf_desc_address(&rtex->resource,
rtex->db_compatible &&
rview->is_stencil_sampler;
- si_set_mutable_tex_desc_fields(rtex,
+ si_set_mutable_tex_desc_fields(sctx->screen, rtex,
rview->base_level_info,
rview->base_level,
rview->base.u.tex.first_level,
desc);
}
- if (rtex->resource.b.b.target != PIPE_BUFFER &&
- rtex->fmask.size) {
+ if (!is_buffer && rtex->fmask.size) {
memcpy(desc + 8,
rview->fmask_state, 8*4);
} else {
if (views->sampler_states[slot])
memcpy(desc + 12,
- views->sampler_states[slot], 4*4);
+ views->sampler_states[slot]->val, 4*4);
}
views->enabled_mask |= 1u << slot;
rview->is_stencil_sampler, true);
} else {
pipe_sampler_view_reference(&views->views[slot], NULL);
- memcpy(descs->list + slot*16, null_texture_descriptor, 8*4);
+ memcpy(desc, null_texture_descriptor, 8*4);
/* Only clear the lower dwords of FMASK. */
- memcpy(descs->list + slot*16 + 8, null_texture_descriptor, 4*4);
+ memcpy(desc + 8, null_texture_descriptor, 4*4);
+ /* Re-set the sampler state if we are transitioning from FMASK. */
+ if (views->sampler_states[slot])
+ memcpy(desc + 12,
+ views->sampler_states[slot]->val, 4*4);
+
views->enabled_mask &= ~(1u << slot);
}
static bool is_compressed_colortex(struct r600_texture *rtex)
{
- return rtex->cmask.size || rtex->fmask.size ||
- (rtex->dcc_offset && rtex->dirty_level_mask);
+ return rtex->fmask.size ||
+ (rtex->dirty_level_mask &&
+ (rtex->cmask.size || rtex->dcc_offset));
+}
+
+static bool depth_needs_decompression(struct r600_texture *rtex,
+ struct si_sampler_view *sview)
+{
+ return rtex->db_compatible &&
+ (!rtex->tc_compatible_htile || sview->is_stencil_sampler);
+}
+
+static void si_update_compressed_tex_shader_mask(struct si_context *sctx,
+ unsigned shader)
+{
+ struct si_textures_info *samplers = &sctx->samplers[shader];
+ unsigned shader_bit = 1 << shader;
+
+ if (samplers->depth_texture_mask ||
+ samplers->compressed_colortex_mask ||
+ sctx->images[shader].compressed_colortex_mask)
+ sctx->compressed_tex_shader_mask |= shader_bit;
+ else
+ sctx->compressed_tex_shader_mask &= ~shader_bit;
}
static void si_set_sampler_views(struct pipe_context *ctx,
(struct r600_texture*)views[i]->texture;
struct si_sampler_view *rview = (struct si_sampler_view *)views[i];
- if (rtex->db_compatible &&
- (!rtex->tc_compatible_htile || rview->is_stencil_sampler)) {
+ if (depth_needs_decompression(rtex, rview)) {
samplers->depth_texture_mask |= 1u << slot;
} else {
samplers->depth_texture_mask &= ~(1u << slot);
samplers->compressed_colortex_mask &= ~(1u << slot);
}
}
+
+ si_update_compressed_tex_shader_mask(sctx, shader);
}
static void
static void si_set_shader_image(struct si_context *ctx,
unsigned shader,
- unsigned slot, const struct pipe_image_view *view)
+ unsigned slot, const struct pipe_image_view *view,
+ bool skip_decompress)
{
struct si_screen *screen = ctx->screen;
struct si_images_info *images = &ctx->images[shader];
si_make_buffer_descriptor(screen, res,
view->format,
view->u.buf.offset,
- view->u.buf.size,
- descs->list + slot * 8);
+ view->u.buf.size, desc);
si_set_buf_desc_address(res, view->u.buf.offset, desc + 4);
images->compressed_colortex_mask &= ~(1 << slot);
static const unsigned char swizzle[4] = { 0, 1, 2, 3 };
struct r600_texture *tex = (struct r600_texture *)res;
unsigned level = view->u.tex.level;
- unsigned width, height, depth;
- bool uses_dcc = tex->dcc_offset &&
- level < tex->surface.num_dcc_levels;
+ unsigned width, height, depth, hw_level;
+ bool uses_dcc = vi_dcc_enabled(tex, level);
assert(!tex->is_depth);
assert(tex->fmask.size == 0);
- if (uses_dcc &&
+ if (uses_dcc && !skip_decompress &&
(view->access & PIPE_IMAGE_ACCESS_WRITE ||
!vi_dcc_formats_compatible(res->b.b.format, view->format))) {
/* If DCC can't be disabled, at least decompress it.
p_atomic_read(&tex->framebuffers_bound))
ctx->need_check_render_feedback = true;
- /* Always force the base level to the selected level.
- *
- * This is required for 3D textures, where otherwise
- * selecting a single slice for non-layered bindings
- * fails. It doesn't hurt the other targets.
- */
- width = u_minify(res->b.b.width0, level);
- height = u_minify(res->b.b.height0, level);
- depth = u_minify(res->b.b.depth0, level);
+ if (ctx->b.chip_class >= GFX9) {
+ /* Always set the base address. The swizzle modes don't
+ * allow setting mipmap level offsets as the base.
+ */
+ width = res->b.b.width0;
+ height = res->b.b.height0;
+ depth = res->b.b.depth0;
+ hw_level = level;
+ } else {
+ /* Always force the base level to the selected level.
+ *
+ * This is required for 3D textures, where otherwise
+ * selecting a single slice for non-layered bindings
+ * fails. It doesn't hurt the other targets.
+ */
+ width = u_minify(res->b.b.width0, level);
+ height = u_minify(res->b.b.height0, level);
+ depth = u_minify(res->b.b.depth0, level);
+ hw_level = 0;
+ }
si_make_texture_descriptor(screen, tex,
false, res->b.b.target,
view->format, swizzle,
- 0, 0,
+ hw_level, hw_level,
view->u.tex.first_layer,
view->u.tex.last_layer,
width, height, depth,
desc, NULL);
- si_set_mutable_tex_desc_fields(tex, &tex->surface.level[level],
+ si_set_mutable_tex_desc_fields(screen, tex,
+ &tex->surface.u.legacy.level[level],
level, level,
util_format_get_blockwidth(view->format),
false, desc);
if (views) {
for (i = 0, slot = start_slot; i < count; ++i, ++slot)
- si_set_shader_image(ctx, shader, slot, &views[i]);
+ si_set_shader_image(ctx, shader, slot, &views[i], false);
} else {
for (i = 0, slot = start_slot; i < count; ++i, ++slot)
- si_set_shader_image(ctx, shader, slot, NULL);
+ si_set_shader_image(ctx, shader, slot, NULL, false);
}
+
+ si_update_compressed_tex_shader_mask(ctx, shader);
}
static void
sstates[i] == samplers->views.sampler_states[slot])
continue;
+#ifdef DEBUG
+ assert(sstates[i]->magic == SI_SAMPLER_STATE_MAGIC);
+#endif
samplers->views.sampler_states[slot] = sstates[i];
/* If FMASK is bound, don't overwrite it.
* The sampler state will be set after FMASK is unbound.
*/
- if (samplers->views.views[i] &&
- samplers->views.views[i]->texture &&
- samplers->views.views[i]->texture->target != PIPE_BUFFER &&
- ((struct r600_texture*)samplers->views.views[i]->texture)->fmask.size)
+ if (samplers->views.views[slot] &&
+ samplers->views.views[slot]->texture &&
+ samplers->views.views[slot]->texture->target != PIPE_BUFFER &&
+ ((struct r600_texture*)samplers->views.views[slot]->texture)->fmask.size)
continue;
memcpy(desc->list + slot * 16 + 12, sstates[i]->val, 4*4);
if (vb >= ARRAY_SIZE(sctx->vertex_buffer))
continue;
- if (!sctx->vertex_buffer[vb].buffer)
+ if (!sctx->vertex_buffer[vb].buffer.resource)
continue;
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
- (struct r600_resource*)sctx->vertex_buffer[vb].buffer,
+ (struct r600_resource*)sctx->vertex_buffer[vb].buffer.resource,
RADEON_USAGE_READ, RADEON_PRIO_VERTEX_BUFFER);
}
bool si_upload_vertex_buffer_descriptors(struct si_context *sctx)
{
+ struct si_vertex_element *velems = sctx->vertex_elements;
struct si_descriptors *desc = &sctx->vertex_buffers;
- bool bound[SI_NUM_VERTEX_BUFFERS] = {};
- unsigned i, count = sctx->vertex_elements->count;
+ unsigned i, count;
+ unsigned desc_list_byte_size;
+ unsigned first_vb_use_mask;
uint64_t va;
uint32_t *ptr;
- if (!sctx->vertex_buffers_dirty)
+ if (!sctx->vertex_buffers_dirty || !velems)
return true;
- if (!count || !sctx->vertex_elements)
+
+ count = velems->count;
+
+ if (!count)
return true;
+ desc_list_byte_size = velems->desc_list_byte_size;
+ first_vb_use_mask = velems->first_vb_use_mask;
+
/* Vertex buffer descriptors are the only ones which are uploaded
* directly through a staging buffer and don't go through
* the fine-grained upload path.
*/
- u_upload_alloc(sctx->b.uploader, 0, count * 16, 256, &desc->buffer_offset,
+ u_upload_alloc(sctx->b.b.const_uploader, 0,
+ desc_list_byte_size,
+ si_optimal_tcc_alignment(sctx, desc_list_byte_size),
+ &desc->buffer_offset,
(struct pipe_resource**)&desc->buffer, (void**)&ptr);
if (!desc->buffer)
return false;
desc->buffer, RADEON_USAGE_READ,
RADEON_PRIO_DESCRIPTORS);
- assert(count <= SI_NUM_VERTEX_BUFFERS);
+ assert(count <= SI_MAX_ATTRIBS);
for (i = 0; i < count; i++) {
- struct pipe_vertex_element *ve = &sctx->vertex_elements->elements[i];
+ struct pipe_vertex_element *ve = &velems->elements[i];
struct pipe_vertex_buffer *vb;
struct r600_resource *rbuffer;
unsigned offset;
+ unsigned vbo_index = ve->vertex_buffer_index;
uint32_t *desc = &ptr[i*4];
- if (ve->vertex_buffer_index >= ARRAY_SIZE(sctx->vertex_buffer)) {
- memset(desc, 0, 16);
- continue;
- }
-
- vb = &sctx->vertex_buffer[ve->vertex_buffer_index];
- rbuffer = (struct r600_resource*)vb->buffer;
+ vb = &sctx->vertex_buffer[vbo_index];
+ rbuffer = (struct r600_resource*)vb->buffer.resource;
if (!rbuffer) {
memset(desc, 0, 16);
continue;
desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32) |
S_008F04_STRIDE(vb->stride);
- if (sctx->b.chip_class <= CIK && vb->stride) {
+ if (sctx->b.chip_class != VI && vb->stride) {
/* Round up by rounding down and adding 1 */
- desc[2] = (vb->buffer->width0 - offset -
- sctx->vertex_elements->format_size[i]) /
+ desc[2] = (vb->buffer.resource->width0 - offset -
+ velems->format_size[i]) /
vb->stride + 1;
} else {
- uint32_t size3;
-
- desc[2] = vb->buffer->width0 - offset;
-
- /* For attributes of size 3 with byte or short
- * components, we use a 4-component data format.
- *
- * As a consequence, we have to round the buffer size
- * up so that the hardware sees four components as
- * being inside the buffer if and only if the first
- * three components are in the buffer.
- *
- * Since the offset and stride are guaranteed to be
- * 4-byte aligned, this alignment will never cross the
- * winsys buffer boundary.
- */
- size3 = (sctx->vertex_elements->fix_size3 >> (2 * i)) & 3;
- if (vb->stride && size3) {
- assert(offset % 4 == 0 && vb->stride % 4 == 0);
- assert(size3 <= 2);
- desc[2] = align(desc[2], size3 * 2);
- }
+ desc[2] = vb->buffer.resource->width0 - offset;
}
- desc[3] = sctx->vertex_elements->rsrc_word3[i];
+ desc[3] = velems->rsrc_word3[i];
- if (!bound[ve->vertex_buffer_index]) {
+ if (first_vb_use_mask & (1 << i)) {
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
- (struct r600_resource*)vb->buffer,
+ (struct r600_resource*)vb->buffer.resource,
RADEON_USAGE_READ, RADEON_PRIO_VERTEX_BUFFER);
- bound[ve->vertex_buffer_index] = true;
}
}
* on performance (confirmed by testing). New descriptors are always
* uploaded to a fresh new buffer, so I don't think flushing the const
* cache is needed. */
- desc->pointer_dirty = true;
si_mark_atom_dirty(sctx, &sctx->shader_userdata.atom);
+ if (sctx->b.chip_class >= CIK)
+ si_mark_atom_dirty(sctx, &sctx->prefetch_L2);
sctx->vertex_buffers_dirty = false;
+ sctx->vertex_buffer_pointer_dirty = true;
return true;
}
{
void *tmp;
- u_upload_alloc(sctx->b.uploader, 0, size, 256, const_offset,
+ u_upload_alloc(sctx->b.b.const_uploader, 0, size,
+ si_optimal_tcc_alignment(sctx, size),
+ const_offset,
(struct pipe_resource**)rbuffer, &tmp);
if (*rbuffer)
util_memcpy_cpu_to_le32(tmp, ptr, size);
}
static void si_pipe_set_constant_buffer(struct pipe_context *ctx,
- uint shader, uint slot,
+ enum pipe_shader_type shader, uint slot,
const struct pipe_constant_buffer *input)
{
struct si_context *sctx = (struct si_context *)ctx;
descs->dirty_mask |= 1u << slot;
sctx->descriptors_dirty |=
1u << si_shader_buffer_descriptors_idx(shader);
+
+ util_range_add(&buf->valid_buffer_range, sbuffer->buffer_offset,
+ sbuffer->buffer_offset + sbuffer->buffer_size);
}
}
-void si_get_shader_buffers(struct si_context *sctx, uint shader,
+void si_get_shader_buffers(struct si_context *sctx,
+ enum pipe_shader_type shader,
uint start_slot, uint count,
struct pipe_shader_buffer *sbuf)
{
S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
- S_008F0C_ELEMENT_SIZE(element_size) |
S_008F0C_INDEX_STRIDE(index_stride) |
S_008F0C_ADD_TID_ENABLE(add_tid);
+ if (sctx->b.chip_class >= GFX9)
+ assert(!swizzle || element_size == 1); /* always 4 bytes on GFX9 */
+ else
+ desc[3] |= S_008F0C_ELEMENT_SIZE(element_size);
+
pipe_resource_reference(&buffers->buffers[slot], buffer);
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
(struct r600_resource*)buffer,
for (int i = 0; i < SI_NUM_SHADERS; ++i) {
si_samplers_update_compressed_colortex_mask(&sctx->samplers[i]);
si_images_update_compressed_colortex_mask(&sctx->images[i]);
+ si_update_compressed_tex_shader_mask(sctx, i);
}
}
}
}
-/* Reallocate a buffer a update all resource bindings where the buffer is
- * bound.
- *
- * This is used to avoid CPU-GPU synchronizations, because it makes the buffer
- * idle by discarding its contents. Apps usually tell us when to do this using
- * map_buffer flags, for example.
- */
-static void si_invalidate_buffer(struct pipe_context *ctx, struct pipe_resource *buf)
+static void si_rebind_buffer(struct pipe_context *ctx, struct pipe_resource *buf,
+ uint64_t old_va)
{
struct si_context *sctx = (struct si_context*)ctx;
struct r600_resource *rbuffer = r600_resource(buf);
unsigned i, shader;
- uint64_t old_va = rbuffer->gpu_address;
unsigned num_elems = sctx->vertex_elements ?
sctx->vertex_elements->count : 0;
- /* Reallocate the buffer in the same pipe_resource. */
- r600_alloc_resource(&sctx->screen->b, rbuffer);
-
/* We changed the buffer, now we need to bind it where the old one
* was bound. This consists of 2 things:
* 1) Updating the resource descriptor and dirtying it.
if (vb >= ARRAY_SIZE(sctx->vertex_buffer))
continue;
- if (!sctx->vertex_buffer[vb].buffer)
+ if (!sctx->vertex_buffer[vb].buffer.resource)
continue;
- if (sctx->vertex_buffer[vb].buffer == buf) {
+ if (sctx->vertex_buffer[vb].buffer.resource == buf) {
sctx->vertex_buffers_dirty = true;
break;
}
}
}
+/* Reallocate a buffer a update all resource bindings where the buffer is
+ * bound.
+ *
+ * This is used to avoid CPU-GPU synchronizations, because it makes the buffer
+ * idle by discarding its contents. Apps usually tell us when to do this using
+ * map_buffer flags, for example.
+ */
+static void si_invalidate_buffer(struct pipe_context *ctx, struct pipe_resource *buf)
+{
+ struct si_context *sctx = (struct si_context*)ctx;
+ struct r600_resource *rbuffer = r600_resource(buf);
+ uint64_t old_va = rbuffer->gpu_address;
+
+ /* Reallocate the buffer in the same pipe_resource. */
+ r600_alloc_resource(&sctx->screen->b, rbuffer);
+
+ si_rebind_buffer(ctx, buf, old_va);
+}
+
/* Update mutable image descriptor fields of all bound textures. */
void si_update_all_texture_descriptors(struct si_context *sctx)
{
view->resource->target == PIPE_BUFFER)
continue;
- si_set_shader_image(sctx, shader, i, view);
+ si_set_shader_image(sctx, shader, i, view, true);
}
/* Sampler views. */
si_set_sampler_view(sctx, shader, i,
samplers->views[i], true);
}
+
+ si_update_compressed_tex_shader_mask(sctx, shader);
}
}
static void si_mark_shader_pointers_dirty(struct si_context *sctx,
unsigned shader)
{
- struct si_descriptors *descs =
- &sctx->descriptors[SI_DESCS_FIRST_SHADER + shader * SI_NUM_SHADER_DESCS];
-
- for (unsigned i = 0; i < SI_NUM_SHADER_DESCS; ++i, ++descs)
- descs->pointer_dirty = true;
+ sctx->shader_pointers_dirty |=
+ u_bit_consecutive(SI_DESCS_FIRST_SHADER + shader * SI_NUM_SHADER_DESCS,
+ SI_NUM_SHADER_DESCS);
if (shader == PIPE_SHADER_VERTEX)
- sctx->vertex_buffers.pointer_dirty = true;
+ sctx->vertex_buffer_pointer_dirty = sctx->vertex_buffers.buffer != NULL;
si_mark_atom_dirty(sctx, &sctx->shader_userdata.atom);
}
static void si_shader_userdata_begin_new_cs(struct si_context *sctx)
{
- int i;
-
- for (i = 0; i < SI_NUM_SHADERS; i++) {
- si_mark_shader_pointers_dirty(sctx, i);
- }
- sctx->descriptors[SI_DESCS_RW_BUFFERS].pointer_dirty = true;
+ sctx->shader_pointers_dirty = u_bit_consecutive(0, SI_NUM_DESCS);
+ sctx->vertex_buffer_pointer_dirty = sctx->vertex_buffers.buffer != NULL;
+ si_mark_atom_dirty(sctx, &sctx->shader_userdata.atom);
}
/* Set a base register address for user data constants in the given shader.
if (*base != new_base) {
*base = new_base;
- if (new_base)
+ if (new_base) {
si_mark_shader_pointers_dirty(sctx, shader);
+
+ if (shader == PIPE_SHADER_VERTEX)
+ sctx->last_vs_state = ~0;
+ }
}
}
void si_shader_change_notify(struct si_context *sctx)
{
/* VS can be bound as VS, ES, or LS. */
- if (sctx->tes_shader.cso)
- si_set_user_data_base(sctx, PIPE_SHADER_VERTEX,
- R_00B530_SPI_SHADER_USER_DATA_LS_0);
- else if (sctx->gs_shader.cso)
+ if (sctx->tes_shader.cso) {
+ if (sctx->b.chip_class >= GFX9) {
+ si_set_user_data_base(sctx, PIPE_SHADER_VERTEX,
+ R_00B430_SPI_SHADER_USER_DATA_LS_0);
+ } else {
+ si_set_user_data_base(sctx, PIPE_SHADER_VERTEX,
+ R_00B530_SPI_SHADER_USER_DATA_LS_0);
+ }
+ } else if (sctx->gs_shader.cso) {
si_set_user_data_base(sctx, PIPE_SHADER_VERTEX,
R_00B330_SPI_SHADER_USER_DATA_ES_0);
- else
+ } else {
si_set_user_data_base(sctx, PIPE_SHADER_VERTEX,
R_00B130_SPI_SHADER_USER_DATA_VS_0);
+ }
/* TES can be bound as ES, VS, or not bound. */
if (sctx->tes_shader.cso) {
static void si_emit_shader_pointer(struct si_context *sctx,
struct si_descriptors *desc,
- unsigned sh_base, bool keep_dirty)
+ unsigned sh_base)
{
struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
uint64_t va;
- if (!desc->pointer_dirty || !desc->buffer)
- return;
+ assert(desc->buffer);
va = desc->buffer->gpu_address +
desc->buffer_offset;
radeon_emit(cs, (sh_base + desc->shader_userdata_offset - SI_SH_REG_OFFSET) >> 2);
radeon_emit(cs, va);
radeon_emit(cs, va >> 32);
-
- desc->pointer_dirty = keep_dirty;
}
void si_emit_graphics_shader_userdata(struct si_context *sctx,
struct r600_atom *atom)
{
- unsigned shader;
+ unsigned mask;
uint32_t *sh_base = sctx->shader_userdata.sh_base;
struct si_descriptors *descs;
descs = &sctx->descriptors[SI_DESCS_RW_BUFFERS];
- if (descs->pointer_dirty) {
- si_emit_shader_pointer(sctx, descs,
- R_00B030_SPI_SHADER_USER_DATA_PS_0, true);
- si_emit_shader_pointer(sctx, descs,
- R_00B130_SPI_SHADER_USER_DATA_VS_0, true);
+ if (sctx->shader_pointers_dirty & (1 << SI_DESCS_RW_BUFFERS)) {
si_emit_shader_pointer(sctx, descs,
- R_00B230_SPI_SHADER_USER_DATA_GS_0, true);
+ R_00B030_SPI_SHADER_USER_DATA_PS_0);
si_emit_shader_pointer(sctx, descs,
- R_00B330_SPI_SHADER_USER_DATA_ES_0, true);
- si_emit_shader_pointer(sctx, descs,
- R_00B430_SPI_SHADER_USER_DATA_HS_0, true);
- descs->pointer_dirty = false;
+ R_00B130_SPI_SHADER_USER_DATA_VS_0);
+
+ if (sctx->b.chip_class >= GFX9) {
+ /* GFX9 merged LS-HS and ES-GS.
+ * Set RW_BUFFERS in the special registers, so that
+ * it's preloaded into s[0:1] instead of s[8:9].
+ */
+ si_emit_shader_pointer(sctx, descs,
+ R_00B208_SPI_SHADER_USER_DATA_ADDR_LO_GS);
+ si_emit_shader_pointer(sctx, descs,
+ R_00B408_SPI_SHADER_USER_DATA_ADDR_LO_HS);
+ } else {
+ si_emit_shader_pointer(sctx, descs,
+ R_00B230_SPI_SHADER_USER_DATA_GS_0);
+ si_emit_shader_pointer(sctx, descs,
+ R_00B330_SPI_SHADER_USER_DATA_ES_0);
+ si_emit_shader_pointer(sctx, descs,
+ R_00B430_SPI_SHADER_USER_DATA_HS_0);
+ }
}
- descs = &sctx->descriptors[SI_DESCS_FIRST_SHADER];
+ mask = sctx->shader_pointers_dirty &
+ u_bit_consecutive(SI_DESCS_FIRST_SHADER,
+ SI_DESCS_FIRST_COMPUTE - SI_DESCS_FIRST_SHADER);
- for (shader = 0; shader < SI_NUM_GRAPHICS_SHADERS; shader++) {
+ while (mask) {
+ unsigned i = u_bit_scan(&mask);
+ unsigned shader = (i - SI_DESCS_FIRST_SHADER) / SI_NUM_SHADER_DESCS;
unsigned base = sh_base[shader];
- unsigned i;
- if (!base)
- continue;
+ if (base)
+ si_emit_shader_pointer(sctx, descs + i, base);
+ }
+ sctx->shader_pointers_dirty &=
+ ~u_bit_consecutive(SI_DESCS_RW_BUFFERS, SI_DESCS_FIRST_COMPUTE);
- for (i = 0; i < SI_NUM_SHADER_DESCS; i++, descs++)
- si_emit_shader_pointer(sctx, descs, base, false);
+ if (sctx->vertex_buffer_pointer_dirty) {
+ si_emit_shader_pointer(sctx, &sctx->vertex_buffers,
+ sh_base[PIPE_SHADER_VERTEX]);
+ sctx->vertex_buffer_pointer_dirty = false;
}
- si_emit_shader_pointer(sctx, &sctx->vertex_buffers, sh_base[PIPE_SHADER_VERTEX], false);
}
void si_emit_compute_shader_userdata(struct si_context *sctx)
{
unsigned base = R_00B900_COMPUTE_USER_DATA_0;
- struct si_descriptors *descs = &sctx->descriptors[SI_DESCS_FIRST_COMPUTE];
+ struct si_descriptors *descs = sctx->descriptors;
+ unsigned compute_mask =
+ u_bit_consecutive(SI_DESCS_FIRST_COMPUTE, SI_NUM_SHADER_DESCS);
+ unsigned mask = sctx->shader_pointers_dirty & compute_mask;
- for (unsigned i = 0; i < SI_NUM_SHADER_DESCS; ++i, ++descs)
- si_emit_shader_pointer(sctx, descs, base, false);
+ while (mask) {
+ unsigned i = u_bit_scan(&mask);
+
+ si_emit_shader_pointer(sctx, descs + i, base);
+ }
+ sctx->shader_pointers_dirty &= ~compute_mask;
}
/* INIT/DEINIT/UPLOAD */
int i;
unsigned ce_offset = 0;
+ STATIC_ASSERT(GFX9_SGPR_TCS_CONST_BUFFERS % 2 == 0);
+ STATIC_ASSERT(GFX9_SGPR_GS_CONST_BUFFERS % 2 == 0);
+
for (i = 0; i < SI_NUM_SHADERS; i++) {
+ bool gfx9_tcs = sctx->b.chip_class == GFX9 &&
+ i == PIPE_SHADER_TESS_CTRL;
+ bool gfx9_gs = sctx->b.chip_class == GFX9 &&
+ i == PIPE_SHADER_GEOMETRY;
+ /* GFX9 has only 4KB of CE, while previous chips had 32KB.
+ * Rarely used descriptors don't use CE RAM.
+ */
+ bool big_ce = sctx->b.chip_class <= VI;
+ bool images_use_ce = big_ce;
+ bool shaderbufs_use_ce = big_ce ||
+ i == PIPE_SHADER_COMPUTE;
+ bool samplers_use_ce = big_ce ||
+ i == PIPE_SHADER_FRAGMENT;
+
si_init_buffer_resources(&sctx->const_buffers[i],
si_const_buffer_descriptors(sctx, i),
- SI_NUM_CONST_BUFFERS, SI_SGPR_CONST_BUFFERS,
+ SI_NUM_CONST_BUFFERS,
+ gfx9_tcs ? GFX9_SGPR_TCS_CONST_BUFFERS :
+ gfx9_gs ? GFX9_SGPR_GS_CONST_BUFFERS :
+ SI_SGPR_CONST_BUFFERS,
RADEON_USAGE_READ, RADEON_PRIO_CONST_BUFFER,
&ce_offset);
si_init_buffer_resources(&sctx->shader_buffers[i],
si_shader_buffer_descriptors(sctx, i),
- SI_NUM_SHADER_BUFFERS, SI_SGPR_SHADER_BUFFERS,
+ SI_NUM_SHADER_BUFFERS,
+ gfx9_tcs ? GFX9_SGPR_TCS_SHADER_BUFFERS :
+ gfx9_gs ? GFX9_SGPR_GS_SHADER_BUFFERS :
+ SI_SGPR_SHADER_BUFFERS,
RADEON_USAGE_READWRITE, RADEON_PRIO_SHADER_RW_BUFFER,
- &ce_offset);
+ shaderbufs_use_ce ? &ce_offset : NULL);
si_init_descriptors(si_sampler_descriptors(sctx, i),
- SI_SGPR_SAMPLERS, 16, SI_NUM_SAMPLERS,
- null_texture_descriptor, &ce_offset);
+ gfx9_tcs ? GFX9_SGPR_TCS_SAMPLERS :
+ gfx9_gs ? GFX9_SGPR_GS_SAMPLERS :
+ SI_SGPR_SAMPLERS,
+ 16, SI_NUM_SAMPLERS,
+ null_texture_descriptor,
+ samplers_use_ce ? &ce_offset : NULL);
si_init_descriptors(si_image_descriptors(sctx, i),
- SI_SGPR_IMAGES, 8, SI_NUM_IMAGES,
- null_image_descriptor, &ce_offset);
+ gfx9_tcs ? GFX9_SGPR_TCS_IMAGES :
+ gfx9_gs ? GFX9_SGPR_GS_IMAGES :
+ SI_SGPR_IMAGES,
+ 8, SI_NUM_IMAGES,
+ null_image_descriptor,
+ images_use_ce ? &ce_offset : NULL);
}
si_init_buffer_resources(&sctx->rw_buffers,
sctx->descriptors_dirty = u_bit_consecutive(0, SI_NUM_DESCS);
- assert(ce_offset <= 32768);
+ if (sctx->b.chip_class >= GFX9)
+ assert(ce_offset <= 4096);
+ else
+ assert(ce_offset <= 32768);
/* Set pipe_context functions. */
sctx->b.b.bind_sampler_states = si_bind_sampler_states;
sctx->b.b.set_sampler_views = si_set_sampler_views;
sctx->b.b.set_stream_output_targets = si_set_streamout_targets;
sctx->b.invalidate_buffer = si_invalidate_buffer;
+ sctx->b.rebind_buffer = si_rebind_buffer;
/* Shader user data. */
si_init_atom(sctx, &sctx->shader_userdata.atom, &sctx->atoms.s.shader_userdata,
/* Set default and immutable mappings. */
si_set_user_data_base(sctx, PIPE_SHADER_VERTEX, R_00B130_SPI_SHADER_USER_DATA_VS_0);
- si_set_user_data_base(sctx, PIPE_SHADER_TESS_CTRL, R_00B430_SPI_SHADER_USER_DATA_HS_0);
- si_set_user_data_base(sctx, PIPE_SHADER_GEOMETRY, R_00B230_SPI_SHADER_USER_DATA_GS_0);
+
+ if (sctx->b.chip_class >= GFX9) {
+ si_set_user_data_base(sctx, PIPE_SHADER_TESS_CTRL,
+ R_00B430_SPI_SHADER_USER_DATA_LS_0);
+ si_set_user_data_base(sctx, PIPE_SHADER_GEOMETRY,
+ R_00B330_SPI_SHADER_USER_DATA_ES_0);
+ } else {
+ si_set_user_data_base(sctx, PIPE_SHADER_TESS_CTRL,
+ R_00B430_SPI_SHADER_USER_DATA_HS_0);
+ si_set_user_data_base(sctx, PIPE_SHADER_GEOMETRY,
+ R_00B230_SPI_SHADER_USER_DATA_GS_0);
+ }
si_set_user_data_base(sctx, PIPE_SHADER_FRAGMENT, R_00B030_SPI_SHADER_USER_DATA_PS_0);
}
const unsigned mask = u_bit_consecutive(0, SI_DESCS_FIRST_COMPUTE);
unsigned dirty = sctx->descriptors_dirty & mask;
+ /* Assume nothing will go wrong: */
+ sctx->shader_pointers_dirty |= dirty;
+
while (dirty) {
unsigned i = u_bit_scan(&dirty);
SI_NUM_DESCS - SI_DESCS_FIRST_COMPUTE);
unsigned dirty = sctx->descriptors_dirty & mask;
+ /* Assume nothing will go wrong: */
+ sctx->shader_pointers_dirty |= dirty;
+
while (dirty) {
unsigned i = u_bit_scan(&dirty);
projects / mesa.git / blobdiff