struct iris_bo *bo = iris_resource_bo(*out_res);
iris_use_pinned_bo(batch, bo, false);
- *out_offset += iris_bo_offset_from_base_address(bo);
+ iris_record_state_size(batch->state_sizes,
+ bo->gtt_offset + *out_offset, size);
- iris_record_state_size(batch->state_sizes, *out_offset, size);
+ *out_offset += iris_bo_offset_from_base_address(bo);
return ptr;
}
static void
flush_before_state_base_change(struct iris_batch *batch)
{
+ const struct gen_device_info *devinfo = &batch->screen->devinfo;
+
/* Flush before emitting STATE_BASE_ADDRESS.
*
* This isn't documented anywhere in the PRM. However, it seems to be
"change STATE_BASE_ADDRESS (flushes)",
PIPE_CONTROL_RENDER_TARGET_FLUSH |
PIPE_CONTROL_DEPTH_CACHE_FLUSH |
- PIPE_CONTROL_DATA_CACHE_FLUSH);
+ PIPE_CONTROL_DATA_CACHE_FLUSH |
+ /* GEN:BUG:1606662791:
+ *
+ * Software must program PIPE_CONTROL command
+ * with "HDC Pipeline Flush" prior to
+ * programming of the below two non-pipeline
+ * state :
+ * * STATE_BASE_ADDRESS
+ * * 3DSTATE_BINDING_TABLE_POOL_ALLOC
+ */
+ ((GEN_GEN == 12 && devinfo->revision == 0 /* A0 */ ?
+ PIPE_CONTROL_FLUSH_HDC : 0)));
}
static void
}
static void
-iris_emit_l3_config(struct iris_batch *batch, const struct gen_l3_config *cfg,
- bool has_slm, bool wants_dc_cache)
+iris_emit_l3_config(struct iris_batch *batch,
+ const struct gen_l3_config *cfg)
{
uint32_t reg_val;
#endif
iris_pack_state(L3_ALLOCATION_REG, ®_val, reg) {
-#if GEN_GEN < 12
- reg.SLMEnable = has_slm;
+#if GEN_GEN < 11
+ reg.SLMEnable = cfg->n[GEN_L3P_SLM] > 0;
#endif
#if GEN_GEN == 11
/* WA_1406697149: Bit 9 "Error Detection Behavior Control" must be set
}
static void
-iris_emit_default_l3_config(struct iris_batch *batch,
- const struct gen_device_info *devinfo,
- bool compute)
+iris_emit_default_l3_config(struct iris_batch *batch, bool compute)
{
+ const struct gen_device_info *devinfo = &batch->screen->devinfo;
bool wants_dc_cache = true;
bool has_slm = compute;
const struct gen_l3_weights w =
gen_get_default_l3_weights(devinfo, wants_dc_cache, has_slm);
const struct gen_l3_config *cfg = gen_get_l3_config(devinfo, w);
- iris_emit_l3_config(batch, cfg, has_slm, wants_dc_cache);
+ iris_emit_l3_config(batch, cfg);
}
-#if GEN_GEN == 9 || GEN_GEN == 10
+#if GEN_GEN == 9
static void
iris_enable_obj_preemption(struct iris_batch *batch, bool enable)
{
emit_pipeline_select(batch, _3D);
- iris_emit_default_l3_config(batch, devinfo, false);
+ iris_emit_default_l3_config(batch, false);
init_state_base_address(batch);
#endif
#if GEN_GEN == 11
- iris_pack_state(GENX(SAMPLER_MODE), ®_val, reg) {
- reg.HeaderlessMessageforPreemptableContexts = 1;
- reg.HeaderlessMessageforPreemptableContextsMask = 1;
- }
- iris_emit_lri(batch, SAMPLER_MODE, reg_val);
+ iris_pack_state(GENX(TCCNTLREG), ®_val, reg) {
+ reg.L3DataPartialWriteMergingEnable = true;
+ reg.ColorZPartialWriteMergingEnable = true;
+ reg.URBPartialWriteMergingEnable = true;
+ reg.TCDisable = true;
+ }
+ iris_emit_lri(batch, TCCNTLREG, reg_val);
- /* Bit 1 must be set in HALF_SLICE_CHICKEN7. */
- iris_pack_state(GENX(HALF_SLICE_CHICKEN7), ®_val, reg) {
- reg.EnabledTexelOffsetPrecisionFix = 1;
- reg.EnabledTexelOffsetPrecisionFixMask = 1;
- }
- iris_emit_lri(batch, HALF_SLICE_CHICKEN7, reg_val);
+ iris_pack_state(GENX(SAMPLER_MODE), ®_val, reg) {
+ reg.HeaderlessMessageforPreemptableContexts = 1;
+ reg.HeaderlessMessageforPreemptableContextsMask = 1;
+ }
+ iris_emit_lri(batch, SAMPLER_MODE, reg_val);
- /* Hardware specification recommends disabling repacking for the
- * compatibility with decompression mechanism in display controller.
- */
- if (devinfo->disable_ccs_repack) {
- iris_pack_state(GENX(CACHE_MODE_0), ®_val, reg) {
- reg.DisableRepackingforCompression = true;
- reg.DisableRepackingforCompressionMask = true;
- }
- iris_emit_lri(batch, CACHE_MODE_0, reg_val);
+ /* Bit 1 must be set in HALF_SLICE_CHICKEN7. */
+ iris_pack_state(GENX(HALF_SLICE_CHICKEN7), ®_val, reg) {
+ reg.EnabledTexelOffsetPrecisionFix = 1;
+ reg.EnabledTexelOffsetPrecisionFixMask = 1;
+ }
+ iris_emit_lri(batch, HALF_SLICE_CHICKEN7, reg_val);
+
+ /* Hardware specification recommends disabling repacking for the
+ * compatibility with decompression mechanism in display controller.
+ */
+ if (devinfo->disable_ccs_repack) {
+ iris_pack_state(GENX(CACHE_MODE_0), ®_val, reg) {
+ reg.DisableRepackingforCompression = true;
+ reg.DisableRepackingforCompressionMask = true;
}
+ iris_emit_lri(batch, CACHE_MODE_0, reg_val);
+ }
- iris_upload_slice_hashing_state(batch);
+ iris_upload_slice_hashing_state(batch);
#endif
/* 3DSTATE_DRAWING_RECTANGLE is non-pipelined, so we want to avoid
iris_emit_cmd(batch, GENX(3DSTATE_POLY_STIPPLE_OFFSET), foo);
iris_alloc_push_constants(batch);
-
-#if GEN_GEN == 10
- /* Gen11+ is enabled for us by the kernel. */
- iris_enable_obj_preemption(batch, true);
-#endif
}
static void
{
UNUSED const struct gen_device_info *devinfo = &batch->screen->devinfo;
+ /* GEN:BUG:1607854226:
+ *
+ * Start with pipeline in 3D mode to set the STATE_BASE_ADDRESS.
+ */
+#if GEN_GEN == 12
+ emit_pipeline_select(batch, _3D);
+#else
emit_pipeline_select(batch, GPGPU);
+#endif
- iris_emit_default_l3_config(batch, devinfo, true);
+ iris_emit_default_l3_config(batch, true);
init_state_base_address(batch);
+#if GEN_GEN == 12
+ emit_pipeline_select(batch, GPGPU);
+#endif
+
#if GEN_GEN == 9
if (devinfo->is_geminilake)
init_glk_barrier_mode(batch, GLK_BARRIER_MODE_GPGPU);
sf.PointWidthSource = state->point_size_per_vertex ? Vertex : State;
sf.PointWidth = state->point_size;
+#if GEN_GEN >= 12
+ sf.DerefBlockSize = PerPolyDerefMode;
+#endif
+
if (state->flatshade_first) {
sf.TriangleFanProvokingVertexSelect = 1;
} else {
return;
struct pipe_resource *res = shs->sampler_table.res;
- shs->sampler_table.offset +=
- iris_bo_offset_from_base_address(iris_resource_bo(res));
+ struct iris_bo *bo = iris_resource_bo(res);
- iris_record_state_size(ice->state.sizes, shs->sampler_table.offset, size);
+ iris_record_state_size(ice->state.sizes,
+ bo->gtt_offset + shs->sampler_table.offset, size);
+
+ shs->sampler_table.offset += iris_bo_offset_from_base_address(bo);
/* Make sure all land in the same BO */
iris_border_color_pool_reserve(ice, IRIS_MAX_TEXTURE_SAMPLERS);
vb.AddressModifyEnable = true;
vb.BufferPitch = buffer->stride;
if (res) {
- vb.BufferSize = res->bo->size - (int) buffer->buffer_offset;
+ vb.BufferSize = res->base.width0 - (int) buffer->buffer_offset;
vb.BufferStartingAddress =
ro_bo(NULL, res->bo->gtt_offset + (int) buffer->buffer_offset);
vb.MOCS = mocs(res->bo, &screen->isl_dev);
iris_populate_vs_key(const struct iris_context *ice,
const struct shader_info *info,
gl_shader_stage last_stage,
- struct brw_vs_prog_key *key)
+ struct iris_vs_prog_key *key)
{
const struct iris_rasterizer_state *cso_rast = ice->state.cso_rast;
if (info->clip_distance_array_size == 0 &&
(info->outputs_written & (VARYING_BIT_POS | VARYING_BIT_CLIP_VERTEX)) &&
last_stage == MESA_SHADER_VERTEX)
- key->nr_userclip_plane_consts = cso_rast->num_clip_plane_consts;
+ key->vue.nr_userclip_plane_consts = cso_rast->num_clip_plane_consts;
}
/**
*/
static void
iris_populate_tcs_key(const struct iris_context *ice,
- struct brw_tcs_prog_key *key)
+ struct iris_tcs_prog_key *key)
{
}
iris_populate_tes_key(const struct iris_context *ice,
const struct shader_info *info,
gl_shader_stage last_stage,
- struct brw_tes_prog_key *key)
+ struct iris_tes_prog_key *key)
{
const struct iris_rasterizer_state *cso_rast = ice->state.cso_rast;
if (info->clip_distance_array_size == 0 &&
(info->outputs_written & (VARYING_BIT_POS | VARYING_BIT_CLIP_VERTEX)) &&
last_stage == MESA_SHADER_TESS_EVAL)
- key->nr_userclip_plane_consts = cso_rast->num_clip_plane_consts;
+ key->vue.nr_userclip_plane_consts = cso_rast->num_clip_plane_consts;
}
/**
iris_populate_gs_key(const struct iris_context *ice,
const struct shader_info *info,
gl_shader_stage last_stage,
- struct brw_gs_prog_key *key)
+ struct iris_gs_prog_key *key)
{
const struct iris_rasterizer_state *cso_rast = ice->state.cso_rast;
if (info->clip_distance_array_size == 0 &&
(info->outputs_written & (VARYING_BIT_POS | VARYING_BIT_CLIP_VERTEX)) &&
last_stage == MESA_SHADER_GEOMETRY)
- key->nr_userclip_plane_consts = cso_rast->num_clip_plane_consts;
+ key->vue.nr_userclip_plane_consts = cso_rast->num_clip_plane_consts;
}
/**
static void
iris_populate_fs_key(const struct iris_context *ice,
const struct shader_info *info,
- struct brw_wm_prog_key *key)
+ struct iris_fs_prog_key *key)
{
struct iris_screen *screen = (void *) ice->ctx.screen;
const struct pipe_framebuffer_state *fb = &ice->state.framebuffer;
static void
iris_populate_cs_key(const struct iris_context *ice,
- struct brw_cs_prog_key *key)
+ struct iris_cs_prog_key *key)
{
}
return iris_bo_offset_from_base_address(res->bo) + shader->assembly.offset;
}
-/* Gen11 workaround table #2056 WABTPPrefetchDisable suggests to disable
- * prefetching of binding tables in A0 and B0 steppings. XXX: Revisit
- * this WA on C0 stepping.
- *
- * TODO: Fill out SamplerCount for prefetching?
- */
-
#define INIT_THREAD_DISPATCH_FIELDS(pkt, prefix, stage) \
pkt.KernelStartPointer = KSP(shader); \
- pkt.BindingTableEntryCount = GEN_GEN == 11 ? 0 : \
- shader->bt.size_bytes / 4; \
+ pkt.BindingTableEntryCount = shader->bt.size_bytes / 4; \
pkt.FloatingPointMode = prog_data->use_alt_mode; \
\
pkt.DispatchGRFStartRegisterForURBData = \
iris_pack_command(GENX(3DSTATE_HS), shader->derived_data, hs) {
INIT_THREAD_DISPATCH_FIELDS(hs, Vertex, MESA_SHADER_TESS_CTRL);
+#if GEN_GEN >= 12
+ /* GEN:BUG:1604578095:
+ *
+ * Hang occurs when the number of max threads is less than 2 times
+ * the number of instance count. The number of max threads must be
+ * more than 2 times the number of instance count.
+ */
+ assert((devinfo->max_tcs_threads / 2) > tcs_prog_data->instances);
+#endif
+
hs.InstanceCount = tcs_prog_data->instances - 1;
hs.MaximumNumberofThreads = devinfo->max_tcs_threads - 1;
hs.IncludeVertexHandles = true;
iris_pack_command(GENX(3DSTATE_PS), ps_state, ps) {
ps.VectorMaskEnable = true;
- // XXX: WABTPPrefetchDisable, see above, drop at C0
- ps.BindingTableEntryCount = GEN_GEN == 11 ? 0 :
- shader->bt.size_bytes / 4;
+ ps.BindingTableEntryCount = shader->bt.size_bytes / 4;
ps.FloatingPointMode = prog_data->use_alt_mode;
ps.MaximumNumberofThreadsPerPSD = 64 - (GEN_GEN == 8 ? 2 : 1);
struct iris_batch *batch,
struct iris_sampler_view *isv)
{
- // XXX: ASTC hacks
enum isl_aux_usage aux_usage =
- iris_resource_texture_aux_usage(ice, isv->res, isv->view.format, 0);
+ iris_resource_texture_aux_usage(ice, isv->res, isv->view.format);
iris_use_pinned_bo(batch, isv->res->bo, false);
iris_use_pinned_bo(batch, iris_resource_bo(isv->surface_state.ref.res), false);
flush_before_state_base_change(batch);
+#if GEN_GEN == 12
+ /* GEN:BUG:1607854226:
+ *
+ * Workaround the non pipelined state not applying in MEDIA/GPGPU pipeline
+ * mode by putting the pipeline temporarily in 3D mode..
+ */
+ if (batch->name == IRIS_BATCH_COMPUTE) {
+ iris_emit_cmd(batch, GENX(PIPELINE_SELECT), sel) {
+ sel.MaskBits = 3;
+ sel.PipelineSelection = _3D;
+ }
+ }
+#endif
+
iris_emit_cmd(batch, GENX(STATE_BASE_ADDRESS), sba) {
sba.SurfaceStateBaseAddressModifyEnable = true;
sba.SurfaceStateBaseAddress = ro_bo(binder->bo, 0);
#endif
}
+#if GEN_GEN == 12
+ /* GEN:BUG:1607854226:
+ *
+ * Put the pipeline back into compute mode.
+ */
+ if (batch->name == IRIS_BATCH_COMPUTE) {
+ iris_emit_cmd(batch, GENX(PIPELINE_SELECT), sel) {
+ sel.MaskBits = 3;
+ sel.PipelineSelection = GPGPU;
+ }
+ }
+#endif
+
flush_after_state_base_change(batch);
batch->last_surface_base_address = binder->bo->gtt_offset;
* cached translations.
*/
uint64_t base_addr = gen_aux_map_get_base(aux_map_ctx);
- assert(base_addr != 0 && ALIGN(base_addr, 32 * 1024) == base_addr);
+ assert(base_addr != 0 && align64(base_addr, 32 * 1024) == base_addr);
iris_load_register_imm64(batch, GENX(GFX_AUX_TABLE_BASE_ADDR_num),
base_addr);
batch->last_aux_map_state = aux_map_state_num;
}
#endif
+struct push_bos {
+ struct {
+ struct iris_address addr;
+ uint32_t length;
+ } buffers[4];
+ int buffer_count;
+ uint32_t max_length;
+};
+
+static void
+setup_constant_buffers(struct iris_context *ice,
+ struct iris_batch *batch,
+ int stage,
+ struct push_bos *push_bos)
+{
+ struct iris_shader_state *shs = &ice->state.shaders[stage];
+ struct iris_compiled_shader *shader = ice->shaders.prog[stage];
+ struct brw_stage_prog_data *prog_data = (void *) shader->prog_data;
+
+ uint32_t push_range_sum = 0;
+
+ int n = 0;
+ for (int i = 0; i < 4; i++) {
+ const struct brw_ubo_range *range = &prog_data->ubo_ranges[i];
+
+ if (range->length == 0)
+ continue;
+
+ push_range_sum += range->length;
+
+ if (range->length > push_bos->max_length)
+ push_bos->max_length = range->length;
+
+ /* Range block is a binding table index, map back to UBO index. */
+ unsigned block_index = iris_bti_to_group_index(
+ &shader->bt, IRIS_SURFACE_GROUP_UBO, range->block);
+ assert(block_index != IRIS_SURFACE_NOT_USED);
+
+ struct pipe_shader_buffer *cbuf = &shs->constbuf[block_index];
+ struct iris_resource *res = (void *) cbuf->buffer;
+
+ assert(cbuf->buffer_offset % 32 == 0);
+
+ push_bos->buffers[n].length = range->length;
+ push_bos->buffers[n].addr =
+ res ? ro_bo(res->bo, range->start * 32 + cbuf->buffer_offset)
+ : ro_bo(batch->screen->workaround_bo, 0);
+ n++;
+ }
+
+ /* From the 3DSTATE_CONSTANT_XS and 3DSTATE_CONSTANT_ALL programming notes:
+ *
+ * "The sum of all four read length fields must be less than or
+ * equal to the size of 64."
+ */
+ assert(push_range_sum <= 64);
+
+ push_bos->buffer_count = n;
+}
+
+static void
+emit_push_constant_packets(struct iris_context *ice,
+ struct iris_batch *batch,
+ int stage,
+ const struct push_bos *push_bos)
+{
+ struct iris_compiled_shader *shader = ice->shaders.prog[stage];
+ struct brw_stage_prog_data *prog_data = (void *) shader->prog_data;
+
+ iris_emit_cmd(batch, GENX(3DSTATE_CONSTANT_VS), pkt) {
+ pkt._3DCommandSubOpcode = push_constant_opcodes[stage];
+ if (prog_data) {
+ /* The Skylake PRM contains the following restriction:
+ *
+ * "The driver must ensure The following case does not occur
+ * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
+ * buffer 3 read length equal to zero committed followed by a
+ * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to
+ * zero committed."
+ *
+ * To avoid this, we program the buffers in the highest slots.
+ * This way, slot 0 is only used if slot 3 is also used.
+ */
+ int n = push_bos->buffer_count;
+ assert(n <= 4);
+ const unsigned shift = 4 - n;
+ for (int i = 0; i < n; i++) {
+ pkt.ConstantBody.ReadLength[i + shift] =
+ push_bos->buffers[i].length;
+ pkt.ConstantBody.Buffer[i + shift] = push_bos->buffers[i].addr;
+ }
+ }
+ }
+}
+
+#if GEN_GEN >= 12
+static void
+emit_push_constant_packet_all(struct iris_context *ice,
+ struct iris_batch *batch,
+ uint32_t shader_mask,
+ const struct push_bos *push_bos)
+{
+ if (!push_bos) {
+ iris_emit_cmd(batch, GENX(3DSTATE_CONSTANT_ALL), pc) {
+ pc.ShaderUpdateEnable = shader_mask;
+ }
+ return;
+ }
+
+ const uint32_t n = push_bos->buffer_count;
+ const uint32_t max_pointers = 4;
+ const uint32_t num_dwords = 2 + 2 * n;
+ uint32_t const_all[2 + 2 * max_pointers];
+ uint32_t *dw = &const_all[0];
+
+ assert(n <= max_pointers);
+ iris_pack_command(GENX(3DSTATE_CONSTANT_ALL), dw, all) {
+ all.DWordLength = num_dwords - 2;
+ all.ShaderUpdateEnable = shader_mask;
+ all.PointerBufferMask = (1 << n) - 1;
+ }
+ dw += 2;
+
+ for (int i = 0; i < n; i++) {
+ _iris_pack_state(batch, GENX(3DSTATE_CONSTANT_ALL_DATA),
+ dw + i * 2, data) {
+ data.PointerToConstantBuffer = push_bos->buffers[i].addr;
+ data.ConstantBufferReadLength = push_bos->buffers[i].length;
+ }
+ }
+ iris_batch_emit(batch, const_all, sizeof(uint32_t) * num_dwords);
+}
+#endif
+
static void
iris_upload_dirty_render_state(struct iris_context *ice,
struct iris_batch *batch,
}
}
+ /* GEN:BUG:1604061319
+ *
+ * 3DSTATE_CONSTANT_* needs to be programmed before BTP_*
+ *
+ * Testing shows that all the 3DSTATE_CONSTANT_XS need to be emitted if
+ * any stage has a dirty binding table.
+ */
+ const bool emit_const_wa = GEN_GEN >= 11 &&
+ (dirty & IRIS_ALL_DIRTY_BINDINGS) != 0;
+
+#if GEN_GEN >= 12
+ uint32_t nobuffer_stages = 0;
+#endif
+
for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
- if (!(dirty & (IRIS_DIRTY_CONSTANTS_VS << stage)))
+ if (!(dirty & (IRIS_DIRTY_CONSTANTS_VS << stage)) &&
+ !emit_const_wa)
continue;
struct iris_shader_state *shs = &ice->state.shaders[stage];
if (shs->sysvals_need_upload)
upload_sysvals(ice, stage);
- struct brw_stage_prog_data *prog_data = (void *) shader->prog_data;
-
- iris_emit_cmd(batch, GENX(3DSTATE_CONSTANT_VS), pkt) {
- pkt._3DCommandSubOpcode = push_constant_opcodes[stage];
- if (prog_data) {
- /* The Skylake PRM contains the following restriction:
- *
- * "The driver must ensure The following case does not occur
- * without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
- * buffer 3 read length equal to zero committed followed by a
- * 3DSTATE_CONSTANT_* with buffer 0 read length not equal to
- * zero committed."
- *
- * To avoid this, we program the buffers in the highest slots.
- * This way, slot 0 is only used if slot 3 is also used.
- */
- int n = 3;
-
- for (int i = 3; i >= 0; i--) {
- const struct brw_ubo_range *range = &prog_data->ubo_ranges[i];
-
- if (range->length == 0)
- continue;
-
- /* Range block is a binding table index, map back to UBO index. */
- unsigned block_index = iris_bti_to_group_index(
- &shader->bt, IRIS_SURFACE_GROUP_UBO, range->block);
- assert(block_index != IRIS_SURFACE_NOT_USED);
+ struct push_bos push_bos = {};
+ setup_constant_buffers(ice, batch, stage, &push_bos);
- struct pipe_shader_buffer *cbuf = &shs->constbuf[block_index];
- struct iris_resource *res = (void *) cbuf->buffer;
-
- assert(cbuf->buffer_offset % 32 == 0);
+#if GEN_GEN >= 12
+ /* If this stage doesn't have any push constants, emit it later in a
+ * single CONSTANT_ALL packet with all the other stages.
+ */
+ if (push_bos.buffer_count == 0) {
+ nobuffer_stages |= 1 << stage;
+ continue;
+ }
- pkt.ConstantBody.ReadLength[n] = range->length;
- pkt.ConstantBody.Buffer[n] =
- res ? ro_bo(res->bo, range->start * 32 + cbuf->buffer_offset)
- : ro_bo(batch->screen->workaround_bo, 0);
- n--;
- }
- }
+ /* The Constant Buffer Read Length field from 3DSTATE_CONSTANT_ALL
+ * contains only 5 bits, so we can only use it for buffers smaller than
+ * 32.
+ */
+ if (push_bos.max_length < 32) {
+ emit_push_constant_packet_all(ice, batch, 1 << stage, &push_bos);
+ continue;
}
+#endif
+ emit_push_constant_packets(ice, batch, stage, &push_bos);
}
+#if GEN_GEN >= 12
+ if (nobuffer_stages)
+ emit_push_constant_packet_all(ice, batch, nobuffer_stages, NULL);
+#endif
+
for (int stage = 0; stage <= MESA_SHADER_FRAGMENT; stage++) {
/* Gen9 requires 3DSTATE_BINDING_TABLE_POINTERS_XS to be re-emitted
* in order to commit constants. TODO: Investigate "Disable Gather
}
if (count) {
+#if GEN_GEN >= 11
+ /* Gen11+ doesn't need the cache workaround below */
+ uint64_t bound = dynamic_bound;
+ while (bound) {
+ const int i = u_bit_scan64(&bound);
+ iris_use_optional_res(batch, genx->vertex_buffers[i].resource,
+ false);
+ }
+#else
/* The VF cache designers cut corners, and made the cache key's
* <VertexBufferIndex, Memory Address> tuple only consider the bottom
* 32 bits of the address. If you have two vertex buffers which get
"workaround: VF cache 32-bit key [VB]",
flush_flags);
}
+#endif
const unsigned vb_dwords = GENX(VERTEX_BUFFER_STATE_length);
iris_use_pinned_bo(batch, bo, false);
}
+#if GEN_GEN < 11
/* The VF cache key only uses 32-bits, see vertex buffer comment above */
uint16_t high_bits = bo->gtt_offset >> 32ull;
if (high_bits != ice->state.last_index_bo_high_bits) {
PIPE_CONTROL_CS_STALL);
ice->state.last_index_bo_high_bits = high_bits;
}
+#endif
}
#define _3DPRIM_END_OFFSET 0x2420
0, NULL, 0, 0);
}
+ /* GEN:BUG:1409226450, Wait for EU to be idle before pipe control which
+ * invalidates the instruction cache
+ */
+ if (GEN_GEN == 12 && (flags & PIPE_CONTROL_INSTRUCTION_INVALIDATE)) {
+ iris_emit_raw_pipe_control(batch,
+ "workaround: CS stall before instruction "
+ "cache invalidate",
+ PIPE_CONTROL_CS_STALL |
+ PIPE_CONTROL_STALL_AT_SCOREBOARD, bo, offset,
+ imm);
+ }
+
if (GEN_GEN == 9 && IS_COMPUTE_PIPELINE(batch) && post_sync_flags) {
/* Project: SKL / Argument: LRI Post Sync Operation [23]
*
PIPE_CONTROL_CS_STALL, bo, offset, imm);
}
- if (GEN_GEN == 10 && (flags & PIPE_CONTROL_RENDER_TARGET_FLUSH)) {
- /* Cannonlake:
- * "Before sending a PIPE_CONTROL command with bit 12 set, SW must issue
- * another PIPE_CONTROL with Render Target Cache Flush Enable (bit 12)
- * = 0 and Pipe Control Flush Enable (bit 7) = 1"
- */
- iris_emit_raw_pipe_control(batch,
- "workaround: PC flush before RT flush",
- PIPE_CONTROL_FLUSH_ENABLE, bo, offset, imm);
- }
-
/* "Flush Types" workarounds ---------------------------------------------
* We do these now because they may add post-sync operations or CS stalls.
*/
}
}
- /* #1130 from Gen10 workarounds page:
- *
- * "Enable Depth Stall on every Post Sync Op if Render target Cache
- * Flush is not enabled in same PIPE CONTROL and Enable Pixel score
- * board stall if Render target cache flush is enabled."
- *
- * Applicable to CNL B0 and C0 steppings only.
- *
- * The wording here is unclear, and this workaround doesn't look anything
- * like the internal bug report recommendations, but leave it be for now...
- */
- if (GEN_GEN == 10) {
- if (flags & PIPE_CONTROL_RENDER_TARGET_FLUSH) {
- flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
- } else if (flags & non_lri_post_sync_flags) {
- flags |= PIPE_CONTROL_DEPTH_STALL;
- }
- }
-
if (flags & PIPE_CONTROL_DEPTH_STALL) {
/* From the PIPE_CONTROL instruction table, bit 13 (Depth Stall Enable):
*
flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
}
+ if (GEN_GEN >= 12 && (flags & PIPE_CONTROL_DEPTH_CACHE_FLUSH)) {
+ /* GEN:BUG:1409600907:
+ *
+ * "PIPE_CONTROL with Depth Stall Enable bit must be set
+ * with any PIPE_CONTROL with Depth Flush Enable bit set.
+ */
+ flags |= PIPE_CONTROL_DEPTH_STALL;
+ }
+
/* Emit --------------------------------------------------------------- */
if (INTEL_DEBUG & DEBUG_PIPE_CONTROL) {
fprintf(stderr,
- " PC [%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%"PRIx64"]: %s\n",
+ " PC [%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%s%"PRIx64"]: %s\n",
(flags & PIPE_CONTROL_FLUSH_ENABLE) ? "PipeCon " : "",
(flags & PIPE_CONTROL_CS_STALL) ? "CS " : "",
(flags & PIPE_CONTROL_STALL_AT_SCOREBOARD) ? "Scoreboard " : "",
(flags & PIPE_CONTROL_WRITE_IMMEDIATE) ? "WriteImm " : "",
(flags & PIPE_CONTROL_WRITE_DEPTH_COUNT) ? "WriteZCount " : "",
(flags & PIPE_CONTROL_WRITE_TIMESTAMP) ? "WriteTimestamp " : "",
+ (flags & PIPE_CONTROL_FLUSH_HDC) ? "HDC " : "",
imm, reason);
}
iris_emit_cmd(batch, GENX(PIPE_CONTROL), pc) {
#if GEN_GEN >= 12
pc.TileCacheFlushEnable = flags & PIPE_CONTROL_TILE_CACHE_FLUSH;
+#endif
+#if GEN_GEN >= 11
+ pc.HDCPipelineFlushEnable = flags & PIPE_CONTROL_FLUSH_HDC;
#endif
pc.LRIPostSyncOperation = NoLRIOperation;
pc.PipeControlFlushEnable = flags & PIPE_CONTROL_FLUSH_ENABLE;