* these fields. However, since we will be growing the BO's live, we
* just set them all to the maximum.
*/
- sba.GeneralStateBufferSize = 0xfffff;
+ sba.GeneralStateBufferSize = 0xfffff;
+ sba.IndirectObjectBufferSize = 0xfffff;
+ if (device->physical->use_softpin) {
+ /* With softpin, we use fixed addresses so we actually know how big
+ * our base addresses are.
+ */
+ sba.DynamicStateBufferSize = DYNAMIC_STATE_POOL_SIZE / 4096;
+ sba.InstructionBufferSize = INSTRUCTION_STATE_POOL_SIZE / 4096;
+ } else {
+ sba.DynamicStateBufferSize = 0xfffff;
+ sba.InstructionBufferSize = 0xfffff;
+ }
sba.GeneralStateBufferSizeModifyEnable = true;
- sba.DynamicStateBufferSize = 0xfffff;
- sba.DynamicStateBufferSizeModifyEnable = true;
- sba.IndirectObjectBufferSize = 0xfffff;
sba.IndirectObjectBufferSizeModifyEnable = true;
- sba.InstructionBufferSize = 0xfffff;
+ sba.DynamicStateBufferSizeModifyEnable = true;
sba.InstructionBuffersizeModifyEnable = true;
# else
/* On gen7, we have upper bounds instead. According to the docs,
{
uint32_t depth_plane =
anv_image_aspect_to_plane(image->aspects, VK_IMAGE_ASPECT_DEPTH_BIT);
- if (image->planes[depth_plane].aux_surface.isl.size_B == 0)
+ if (image->planes[depth_plane].aux_usage == ISL_AUX_USAGE_NONE)
return;
const enum isl_aux_state initial_state =
}
}
+#define READ_ONCE(x) (*(volatile __typeof__(x) *)&(x))
+
+#if GEN_GEN == 12
+static void
+anv_image_init_aux_tt(struct anv_cmd_buffer *cmd_buffer,
+ const struct anv_image *image,
+ VkImageAspectFlagBits aspect,
+ uint32_t base_level, uint32_t level_count,
+ uint32_t base_layer, uint32_t layer_count)
+{
+ uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);
+
+ uint64_t base_address =
+ anv_address_physical(image->planes[plane].address);
+
+ const struct isl_surf *isl_surf = &image->planes[plane].surface.isl;
+ uint64_t format_bits = gen_aux_map_format_bits_for_isl_surf(isl_surf);
+
+ /* We're about to live-update the AUX-TT. We really don't want anyone else
+ * trying to read it while we're doing this. We could probably get away
+ * with not having this stall in some cases if we were really careful but
+ * it's better to play it safe. Full stall the GPU.
+ */
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
+ for (uint32_t a = 0; a < layer_count; a++) {
+ const uint32_t layer = base_layer + a;
+
+ uint64_t start_offset_B = UINT64_MAX, end_offset_B = 0;
+ for (uint32_t l = 0; l < level_count; l++) {
+ const uint32_t level = base_level + l;
+
+ uint32_t logical_array_layer, logical_z_offset_px;
+ if (image->type == VK_IMAGE_TYPE_3D) {
+ logical_array_layer = 0;
+
+ /* If the given miplevel does not have this layer, then any higher
+ * miplevels won't either because miplevels only get smaller the
+ * higher the LOD.
+ */
+ assert(layer < image->extent.depth);
+ if (layer >= anv_minify(image->extent.depth, level))
+ break;
+ logical_z_offset_px = layer;
+ } else {
+ assert(layer < image->array_size);
+ logical_array_layer = layer;
+ logical_z_offset_px = 0;
+ }
+
+ uint32_t slice_start_offset_B, slice_end_offset_B;
+ isl_surf_get_image_range_B_tile(isl_surf, level,
+ logical_array_layer,
+ logical_z_offset_px,
+ &slice_start_offset_B,
+ &slice_end_offset_B);
+
+ start_offset_B = MIN2(start_offset_B, slice_start_offset_B);
+ end_offset_B = MAX2(end_offset_B, slice_end_offset_B);
+ }
+
+ /* Aux operates 64K at a time */
+ start_offset_B = align_down_u64(start_offset_B, 64 * 1024);
+ end_offset_B = align_u64(end_offset_B, 64 * 1024);
+
+ for (uint64_t offset = start_offset_B;
+ offset < end_offset_B; offset += 64 * 1024) {
+ uint64_t address = base_address + offset;
+
+ uint64_t aux_entry_addr64, *aux_entry_map;
+ aux_entry_map = gen_aux_map_get_entry(cmd_buffer->device->aux_map_ctx,
+ address, &aux_entry_addr64);
+
+ assert(cmd_buffer->device->physical->use_softpin);
+ struct anv_address aux_entry_address = {
+ .bo = NULL,
+ .offset = aux_entry_addr64,
+ };
+
+ const uint64_t old_aux_entry = READ_ONCE(*aux_entry_map);
+ uint64_t new_aux_entry =
+ (old_aux_entry & GEN_AUX_MAP_ADDRESS_MASK) | format_bits;
+
+ if (isl_aux_usage_has_ccs(image->planes[plane].aux_usage))
+ new_aux_entry |= GEN_AUX_MAP_ENTRY_VALID_BIT;
+
+ gen_mi_store(&b, gen_mi_mem64(aux_entry_address),
+ gen_mi_imm(new_aux_entry));
+ }
+ }
+
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_AUX_TABLE_INVALIDATE_BIT;
+}
+#endif /* GEN_GEN == 12 */
+
/**
* @brief Transitions a color buffer from one layout to another.
*
VkImageLayout initial_layout,
VkImageLayout final_layout)
{
- const struct gen_device_info *devinfo = &cmd_buffer->device->info;
+ struct anv_device *device = cmd_buffer->device;
+ const struct gen_device_info *devinfo = &device->info;
/* Validate the inputs. */
assert(cmd_buffer);
assert(image && image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV);
if (initial_layout == VK_IMAGE_LAYOUT_UNDEFINED ||
initial_layout == VK_IMAGE_LAYOUT_PREINITIALIZED) {
+#if GEN_GEN == 12
+ if (device->physical->has_implicit_ccs && devinfo->has_aux_map) {
+ anv_image_init_aux_tt(cmd_buffer, image, aspect,
+ base_level, level_count,
+ base_layer, layer_count);
+ }
+#else
+ assert(!(device->physical->has_implicit_ccs && devinfo->has_aux_map));
+#endif
+
/* A subresource in the undefined layout may have been aliased and
* populated with any arrangement of bits. Therefore, we must initialize
* the related aux buffer and clear buffer entry with desirable values.
*/
cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_VF_CACHE_INVALIDATE_BIT;
+ /* Re-emit the aux table register in every command buffer. This way we're
+ * ensured that we have the table even if this command buffer doesn't
+ * initialize any images.
+ */
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_AUX_TABLE_INVALIDATE_BIT;
+
/* We send an "Indirect State Pointers Disable" packet at
* EndCommandBuffer, so all push contant packets are ignored during a
* context restore. Documentation says after that command, we need to
uint32_t l3cr;
anv_pack_struct(&l3cr, L3_ALLOCATION_REG,
-#if GEN_GEN < 12
+#if GEN_GEN < 11
.SLMEnable = has_slm,
#endif
#if GEN_GEN == 11
void
genX(cmd_buffer_apply_pipe_flushes)(struct anv_cmd_buffer *cmd_buffer)
{
+ UNUSED const struct gen_device_info *devinfo = &cmd_buffer->device->info;
enum anv_pipe_bits bits = cmd_buffer->state.pending_pipe_bits;
if (cmd_buffer->device->physical->always_flush_cache)
bits |= ANV_PIPE_TILE_CACHE_FLUSH_BIT;
}
+ /* GEN:BUG:1409226450, Wait for EU to be idle before pipe control which
+ * invalidates the instruction cache
+ */
+ if (GEN_GEN == 12 && (bits & ANV_PIPE_INSTRUCTION_CACHE_INVALIDATE_BIT))
+ bits |= ANV_PIPE_CS_STALL_BIT | ANV_PIPE_STALL_AT_SCOREBOARD_BIT;
+
if ((GEN_GEN >= 8 && GEN_GEN <= 9) &&
(bits & ANV_PIPE_CS_STALL_BIT) &&
(bits & ANV_PIPE_VF_CACHE_INVALIDATE_BIT)) {
sizeof(cmd_buffer->state.gfx.ib_dirty_range));
}
+ /* Project: SKL / Argument: LRI Post Sync Operation [23]
+ *
+ * "PIPECONTROL command with “Command Streamer Stall Enable” must be
+ * programmed prior to programming a PIPECONTROL command with "LRI
+ * Post Sync Operation" in GPGPU mode of operation (i.e when
+ * PIPELINE_SELECT command is set to GPGPU mode of operation)."
+ *
+ * The same text exists a few rows below for Post Sync Op.
+ *
+ * On Gen12 this is GEN:BUG:1607156449.
+ */
+ if (bits & ANV_PIPE_POST_SYNC_BIT) {
+ if ((GEN_GEN == 9 || (GEN_GEN == 12 && devinfo->revision == 0 /* A0 */)) &&
+ cmd_buffer->state.current_pipeline == GPGPU)
+ bits |= ANV_PIPE_CS_STALL_BIT;
+ bits &= ~ANV_PIPE_POST_SYNC_BIT;
+ }
+
if (bits & (ANV_PIPE_FLUSH_BITS | ANV_PIPE_CS_STALL_BIT)) {
anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
#if GEN_GEN >= 12
* I chose "Stall at Pixel Scoreboard" since that's what we use in
* mesa and it seems to work fine. The choice is fairly arbitrary.
*/
- if ((bits & ANV_PIPE_CS_STALL_BIT) &&
- !(bits & (ANV_PIPE_FLUSH_BITS | ANV_PIPE_DEPTH_STALL_BIT |
- ANV_PIPE_STALL_AT_SCOREBOARD_BIT)))
+ if (pipe.CommandStreamerStallEnable &&
+ !pipe.RenderTargetCacheFlushEnable &&
+ !pipe.DepthCacheFlushEnable &&
+ !pipe.StallAtPixelScoreboard &&
+ !pipe.PostSyncOperation &&
+ !pipe.DepthStallEnable &&
+ !pipe.DCFlushEnable)
pipe.StallAtPixelScoreboard = true;
}
}
}
+#if GEN_GEN == 12
+ if ((bits & ANV_PIPE_AUX_TABLE_INVALIDATE_BIT) &&
+ cmd_buffer->device->info.has_aux_map) {
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
+ lri.RegisterOffset = GENX(GFX_CCS_AUX_INV_num);
+ lri.DataDWord = 1;
+ }
+ }
+#endif
+
bits &= ~ANV_PIPE_INVALIDATE_BITS;
}
}
}
+static struct anv_cmd_pipeline_state *
+pipe_state_for_stage(struct anv_cmd_buffer *cmd_buffer,
+ gl_shader_stage stage)
+{
+ switch (stage) {
+ case MESA_SHADER_COMPUTE:
+ return &cmd_buffer->state.compute.base;
+
+ case MESA_SHADER_VERTEX:
+ case MESA_SHADER_TESS_CTRL:
+ case MESA_SHADER_TESS_EVAL:
+ case MESA_SHADER_GEOMETRY:
+ case MESA_SHADER_FRAGMENT:
+ return &cmd_buffer->state.gfx.base;
+
+ default:
+ unreachable("invalid stage");
+ }
+}
+
static VkResult
emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
gl_shader_stage stage,
struct anv_state *bt_state)
{
struct anv_subpass *subpass = cmd_buffer->state.subpass;
- struct anv_cmd_pipeline_state *pipe_state;
- struct anv_pipeline *pipeline;
uint32_t state_offset;
- switch (stage) {
- case MESA_SHADER_COMPUTE:
- pipe_state = &cmd_buffer->state.compute.base;
- break;
- default:
- pipe_state = &cmd_buffer->state.gfx.base;
- break;
- }
- pipeline = pipe_state->pipeline;
+ struct anv_cmd_pipeline_state *pipe_state =
+ pipe_state_for_stage(cmd_buffer, stage);
+ struct anv_pipeline *pipeline = pipe_state->pipeline;
if (!anv_pipeline_has_stage(pipeline, stage)) {
*bt_state = (struct anv_state) { 0, };
struct anv_state *state)
{
struct anv_cmd_pipeline_state *pipe_state =
- stage == MESA_SHADER_COMPUTE ? &cmd_buffer->state.compute.base :
- &cmd_buffer->state.gfx.base;
+ pipe_state_for_stage(cmd_buffer, stage);
struct anv_pipeline *pipeline = pipe_state->pipeline;
if (!anv_pipeline_has_stage(pipeline, stage)) {
const struct anv_pipeline_bind_map *bind_map =
&pipeline->shaders[stage]->bind_map;
+#if GEN_GEN >= 12
+ c.MOCS = cmd_buffer->device->isl_dev.mocs.internal;
+#endif
+
#if GEN_GEN >= 8 || GEN_IS_HASWELL
/* The Skylake PRM contains the following restriction:
*
if (count == 0) {
anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_CONSTANT_ALL), c) {
c.ShaderUpdateEnable = shader_mask;
+ c.MOCS = cmd_buffer->device->isl_dev.mocs.internal;
}
return;
}
dw = anv_batch_emitn(&cmd_buffer->batch, num_dwords,
GENX(3DSTATE_CONSTANT_ALL),
.ShaderUpdateEnable = shader_mask,
- .PointerBufferMask = buffers);
+ .PointerBufferMask = buffers,
+ .MOCS = cmd_buffer->device->isl_dev.mocs.internal);
for (int i = 0; i < count; i++) {
const struct anv_push_range *range = &bind_map->push_ranges[i];
cmd_buffer->state.push_constants_dirty &= ~flushed;
}
-#if GEN_GEN >= 12
-void
-genX(cmd_buffer_aux_map_state)(struct anv_cmd_buffer *cmd_buffer)
-{
- void *aux_map_ctx = cmd_buffer->device->aux_map_ctx;
- if (!aux_map_ctx)
- return;
- uint32_t aux_map_state_num = gen_aux_map_get_state_num(aux_map_ctx);
- if (cmd_buffer->state.last_aux_map_state != aux_map_state_num) {
- /* If the aux-map state number increased, then we need to rewrite the
- * register. Rewriting the register is used to both set the aux-map
- * translation table address, and also to invalidate any previously
- * cached translations.
- */
- uint64_t base_addr = gen_aux_map_get_base(aux_map_ctx);
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
- lri.RegisterOffset = GENX(GFX_AUX_TABLE_BASE_ADDR_num);
- lri.DataDWord = base_addr & 0xffffffff;
- }
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
- lri.RegisterOffset = GENX(GFX_AUX_TABLE_BASE_ADDR_num) + 4;
- lri.DataDWord = base_addr >> 32;
- }
- cmd_buffer->state.last_aux_map_state = aux_map_state_num;
- }
-}
-#endif
-
void
genX(cmd_buffer_flush_state)(struct anv_cmd_buffer *cmd_buffer)
{
assert((pipeline->active_stages & VK_SHADER_STAGE_COMPUTE_BIT) == 0);
- genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->urb.l3_config);
+ genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->l3_config);
genX(cmd_buffer_emit_hashing_mode)(cmd_buffer, UINT_MAX, UINT_MAX, 1);
genX(flush_pipeline_select_3d)(cmd_buffer);
-#if GEN_GEN >= 12
- genX(cmd_buffer_aux_map_state)(cmd_buffer);
-#endif
-
if (vb_emit) {
const uint32_t num_buffers = __builtin_popcount(vb_emit);
const uint32_t num_dwords = 1 + num_buffers * 4;
assert(pipeline->active_stages == VK_SHADER_STAGE_COMPUTE_BIT);
- genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->urb.l3_config);
+ genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->l3_config);
genX(flush_pipeline_select_gpgpu)(cmd_buffer);
-#if GEN_GEN >= 12
- genX(cmd_buffer_aux_map_state)(cmd_buffer);
-#endif
-
if (cmd_buffer->state.compute.pipeline_dirty) {
/* From the Sky Lake PRM Vol 2a, MEDIA_VFE_STATE:
*
isl_emit_depth_stencil_hiz_s(&device->isl_dev, dw, &info);
if (GEN_GEN >= 12) {
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
/* GEN:BUG:1408224581
*
* Workaround: Gen12LP Astep only An additional pipe control with
(att_state->fast_clear && !att_state->clear_color_is_zero_one) ||
att_state->input_aux_usage != att_state->aux_usage;
- VkImageLayout target_layout, target_stencil_layout;
+ VkImageLayout target_layout;
if (iview->aspect_mask & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV &&
!input_needs_resolve) {
/* Layout transitions before the final only help to enable sampling
target_layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
} else {
target_layout = subpass->attachments[i].layout;
- target_stencil_layout = subpass->attachments[i].stencil_layout;
}
+ VkImageLayout target_stencil_layout =
+ subpass->attachments[i].stencil_layout;
+
uint32_t base_layer, layer_count;
if (image->type == VK_IMAGE_TYPE_3D) {
base_layer = 0;
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_event, event, _event);
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
if (stageMask & ANV_PIPELINE_STAGE_PIPELINED_BITS) {
pc.StallAtPixelScoreboard = true;
ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
ANV_FROM_HANDLE(anv_event, event, _event);
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_POST_SYNC_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
if (stageMask & ANV_PIPELINE_STAGE_PIPELINED_BITS) {
pc.StallAtPixelScoreboard = true;
projects / mesa.git / blobdiff