#include "common/gen_l3_config.h"
#include "common/gen_sample_positions.h"
+#include "vk_util.h"
#include "vk_format_info.h"
static uint32_t
default: unreachable("Invalid component");
}
+ /*
+ * Take in account hardware restrictions when dealing with 64-bit floats.
+ *
+ * From Broadwell spec, command reference structures, page 586:
+ * "When SourceElementFormat is set to one of the *64*_PASSTHRU formats,
+ * 64-bit components are stored * in the URB without any conversion. In
+ * this case, vertex elements must be written as 128 or 256 bits, with
+ * VFCOMP_STORE_0 being used to pad the output as required. E.g., if
+ * R64_PASSTHRU is used to copy a 64-bit Red component into the URB,
+ * Component 1 must be specified as VFCOMP_STORE_0 (with Components 2,3
+ * set to VFCOMP_NOSTORE) in order to output a 128-bit vertex element, or
+ * Components 1-3 must be specified as VFCOMP_STORE_0 in order to output
+ * a 256-bit vertex element. Likewise, use of R64G64B64_PASSTHRU requires
+ * Component 3 to be specified as VFCOMP_STORE_0 in order to output a
+ * 256-bit vertex element."
+ */
if (bits) {
return VFCOMP_STORE_SRC;
- } else if (comp < 3) {
+ } else if (comp >= 2 &&
+ !isl_format_layouts[format].channels.b.bits &&
+ isl_format_layouts[format].channels.r.type == ISL_RAW) {
+ /* When emitting 64-bit attributes, we need to write either 128 or 256
+ * bit chunks, using VFCOMP_NOSTORE when not writing the chunk, and
+ * VFCOMP_STORE_0 to pad the written chunk */
+ return VFCOMP_NOSTORE;
+ } else if (comp < 3 ||
+ isl_format_layouts[format].channels.r.type == ISL_RAW) {
+ /* Note we need to pad with value 0, not 1, due hardware restrictions
+ * (see comment above) */
return VFCOMP_STORE_0;
} else if (isl_format_layouts[format].channels.r.type == ISL_UINT ||
isl_format_layouts[format].channels.r.type == ISL_SINT) {
/* Pull inputs_read out of the VS prog data */
const uint64_t inputs_read = vs_prog_data->inputs_read;
+ const uint64_t double_inputs_read = vs_prog_data->double_inputs_read;
assert((inputs_read & ((1 << VERT_ATTRIB_GENERIC0) - 1)) == 0);
const uint32_t elements = inputs_read >> VERT_ATTRIB_GENERIC0;
-
-#if GEN_GEN >= 8
- /* On BDW+, we only need to allocate space for base ids. Setting up
- * the actual vertex and instance id is a separate packet.
- */
- const bool needs_svgs_elem = vs_prog_data->uses_basevertex ||
- vs_prog_data->uses_baseinstance;
-#else
- /* On Haswell and prior, vertex and instance id are created by using the
- * ComponentControl fields, so we need an element for any of them.
- */
+ const uint32_t elements_double = double_inputs_read >> VERT_ATTRIB_GENERIC0;
const bool needs_svgs_elem = vs_prog_data->uses_vertexid ||
vs_prog_data->uses_instanceid ||
- vs_prog_data->uses_basevertex ||
+ vs_prog_data->uses_firstvertex ||
vs_prog_data->uses_baseinstance;
-#endif
- uint32_t elem_count = __builtin_popcount(elements) + needs_svgs_elem;
- if (elem_count == 0)
+ uint32_t elem_count = __builtin_popcount(elements) -
+ __builtin_popcount(elements_double) / 2;
+
+ const uint32_t total_elems =
+ elem_count + needs_svgs_elem + vs_prog_data->uses_drawid;
+ if (total_elems == 0)
return;
uint32_t *p;
- const uint32_t num_dwords = 1 + elem_count * 2;
+ const uint32_t num_dwords = 1 + total_elems * 2;
p = anv_batch_emitn(&pipeline->batch, num_dwords,
GENX(3DSTATE_VERTEX_ELEMENTS));
+ if (!p)
+ return;
memset(p + 1, 0, (num_dwords - 1) * 4);
for (uint32_t i = 0; i < info->vertexAttributeDescriptionCount; i++) {
VK_IMAGE_ASPECT_COLOR_BIT,
VK_IMAGE_TILING_LINEAR);
- assert(desc->binding < 32);
+ assert(desc->binding < MAX_VBS);
if ((elements & (1 << desc->location)) == 0)
continue; /* Binding unused */
- uint32_t slot = __builtin_popcount(elements & ((1 << desc->location) - 1));
+ uint32_t slot =
+ __builtin_popcount(elements & ((1 << desc->location) - 1)) -
+ DIV_ROUND_UP(__builtin_popcount(elements_double &
+ ((1 << desc->location) -1)), 2);
struct GENX(VERTEX_ELEMENT_STATE) element = {
.VertexBufferIndex = desc->binding,
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VF_INSTANCING), vfi) {
vfi.InstancingEnable = pipeline->instancing_enable[desc->binding];
vfi.VertexElementIndex = slot;
- /* Vulkan so far doesn't have an instance divisor, so
- * this is always 1 (ignored if not instancing). */
- vfi.InstanceDataStepRate = 1;
+ /* Our implementation of VK_KHR_multiview uses instancing to draw
+ * the different views. If the client asks for instancing, we
+ * need to use the Instance Data Step Rate to ensure that we
+ * repeat the client's per-instance data once for each view.
+ */
+ vfi.InstanceDataStepRate = anv_subpass_view_count(pipeline->subpass);
}
#endif
}
- const uint32_t id_slot = __builtin_popcount(elements);
+ const uint32_t id_slot = elem_count;
if (needs_svgs_elem) {
/* From the Broadwell PRM for the 3D_Vertex_Component_Control enum:
* "Within a VERTEX_ELEMENT_STATE structure, if a Component
* This means, that if we have BaseInstance, we need BaseVertex as
* well. Just do all or nothing.
*/
- uint32_t base_ctrl = (vs_prog_data->uses_basevertex ||
+ uint32_t base_ctrl = (vs_prog_data->uses_firstvertex ||
vs_prog_data->uses_baseinstance) ?
VFCOMP_STORE_SRC : VFCOMP_STORE_0;
struct GENX(VERTEX_ELEMENT_STATE) element = {
- .VertexBufferIndex = 32, /* Reserved for this */
+ .VertexBufferIndex = ANV_SVGS_VB_INDEX,
.Valid = true,
.SourceElementFormat = ISL_FORMAT_R32G32_UINT,
.Component0Control = base_ctrl,
sgvs.InstanceIDElementOffset = id_slot;
}
#endif
+
+ const uint32_t drawid_slot = elem_count + needs_svgs_elem;
+ if (vs_prog_data->uses_drawid) {
+ struct GENX(VERTEX_ELEMENT_STATE) element = {
+ .VertexBufferIndex = ANV_DRAWID_VB_INDEX,
+ .Valid = true,
+ .SourceElementFormat = ISL_FORMAT_R32_UINT,
+ .Component0Control = VFCOMP_STORE_SRC,
+ .Component1Control = VFCOMP_STORE_0,
+ .Component2Control = VFCOMP_STORE_0,
+ .Component3Control = VFCOMP_STORE_0,
+ };
+ GENX(VERTEX_ELEMENT_STATE_pack)(NULL,
+ &p[1 + drawid_slot * 2],
+ &element);
+
+#if GEN_GEN >= 8
+ anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VF_INSTANCING), vfi) {
+ vfi.VertexElementIndex = drawid_slot;
+ }
+#endif
+ }
}
void
genX(emit_urb_setup)(struct anv_device *device, struct anv_batch *batch,
+ const struct gen_l3_config *l3_config,
VkShaderStageFlags active_stages,
- unsigned vs_size, unsigned gs_size,
- const struct gen_l3_config *l3_config)
+ const unsigned entry_size[4])
{
- if (!(active_stages & VK_SHADER_STAGE_VERTEX_BIT))
- vs_size = 1;
-
- if (!(active_stages & VK_SHADER_STAGE_GEOMETRY_BIT))
- gs_size = 1;
-
- unsigned vs_entry_size_bytes = vs_size * 64;
- unsigned gs_entry_size_bytes = gs_size * 64;
-
- /* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
- *
- * VS Number of URB Entries must be divisible by 8 if the VS URB Entry
- * Allocation Size is less than 9 512-bit URB entries.
- *
- * Similar text exists for GS.
- */
- unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
- unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
-
- /* URB allocations must be done in 8k chunks. */
- unsigned chunk_size_bytes = 8192;
-
- /* Determine the size of the URB in chunks. */
- const unsigned total_urb_size =
- gen_get_l3_config_urb_size(&device->info, l3_config);
- const unsigned urb_chunks = total_urb_size * 1024 / chunk_size_bytes;
-
- /* Reserve space for push constants */
- unsigned push_constant_kb;
- if (device->info.gen >= 8)
- push_constant_kb = 32;
- else if (device->info.is_haswell)
- push_constant_kb = device->info.gt == 3 ? 32 : 16;
- else
- push_constant_kb = 16;
-
- unsigned push_constant_bytes = push_constant_kb * 1024;
- unsigned push_constant_chunks =
- push_constant_bytes / chunk_size_bytes;
-
- /* Initially, assign each stage the minimum amount of URB space it needs,
- * and make a note of how much additional space it "wants" (the amount of
- * additional space it could actually make use of).
- */
-
- /* VS has a lower limit on the number of URB entries */
- unsigned vs_chunks =
- ALIGN(device->info.urb.min_vs_entries * vs_entry_size_bytes,
- chunk_size_bytes) / chunk_size_bytes;
- unsigned vs_wants =
- ALIGN(device->info.urb.max_vs_entries * vs_entry_size_bytes,
- chunk_size_bytes) / chunk_size_bytes - vs_chunks;
-
- unsigned gs_chunks = 0;
- unsigned gs_wants = 0;
- if (active_stages & VK_SHADER_STAGE_GEOMETRY_BIT) {
- /* There are two constraints on the minimum amount of URB space we can
- * allocate:
- *
- * (1) We need room for at least 2 URB entries, since we always operate
- * the GS in DUAL_OBJECT mode.
- *
- * (2) We can't allocate less than nr_gs_entries_granularity.
- */
- gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
- chunk_size_bytes) / chunk_size_bytes;
- gs_wants =
- ALIGN(device->info.urb.max_gs_entries * gs_entry_size_bytes,
- chunk_size_bytes) / chunk_size_bytes - gs_chunks;
- }
-
- /* There should always be enough URB space to satisfy the minimum
- * requirements of each stage.
- */
- unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
- assert(total_needs <= urb_chunks);
-
- /* Mete out remaining space (if any) in proportion to "wants". */
- unsigned total_wants = vs_wants + gs_wants;
- unsigned remaining_space = urb_chunks - total_needs;
- if (remaining_space > total_wants)
- remaining_space = total_wants;
- if (remaining_space > 0) {
- unsigned vs_additional = (unsigned)
- round(vs_wants * (((double) remaining_space) / total_wants));
- vs_chunks += vs_additional;
- remaining_space -= vs_additional;
- gs_chunks += remaining_space;
- }
-
- /* Sanity check that we haven't over-allocated. */
- assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
-
- /* Finally, compute the number of entries that can fit in the space
- * allocated to each stage.
- */
- unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
- unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
-
- /* Since we rounded up when computing *_wants, this may be slightly more
- * than the maximum allowed amount, so correct for that.
- */
- nr_vs_entries = MIN2(nr_vs_entries, device->info.urb.max_vs_entries);
- nr_gs_entries = MIN2(nr_gs_entries, device->info.urb.max_gs_entries);
+ const struct gen_device_info *devinfo = &device->info;
+#if GEN_IS_HASWELL
+ const unsigned push_constant_kb = devinfo->gt == 3 ? 32 : 16;
+#else
+ const unsigned push_constant_kb = GEN_GEN >= 8 ? 32 : 16;
+#endif
- /* Ensure that we program a multiple of the granularity. */
- nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
- nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
+ const unsigned urb_size_kb = gen_get_l3_config_urb_size(devinfo, l3_config);
- /* Finally, sanity check to make sure we have at least the minimum number
- * of entries needed for each stage.
- */
- assert(nr_vs_entries >= device->info.urb.min_vs_entries);
- if (active_stages & VK_SHADER_STAGE_GEOMETRY_BIT)
- assert(nr_gs_entries >= 2);
+ unsigned entries[4];
+ unsigned start[4];
+ gen_get_urb_config(devinfo,
+ 1024 * push_constant_kb, 1024 * urb_size_kb,
+ active_stages &
+ VK_SHADER_STAGE_TESSELLATION_EVALUATION_BIT,
+ active_stages & VK_SHADER_STAGE_GEOMETRY_BIT,
+ entry_size, entries, start);
#if GEN_GEN == 7 && !GEN_IS_HASWELL
/* From the IVB PRM Vol. 2, Part 1, Section 3.2.1:
}
#endif
- /* Lay out the URB in the following order:
- * - push constants
- * - VS
- * - GS
- */
- anv_batch_emit(batch, GENX(3DSTATE_URB_VS), urb) {
- urb.VSURBStartingAddress = push_constant_chunks;
- urb.VSURBEntryAllocationSize = vs_size - 1;
- urb.VSNumberofURBEntries = nr_vs_entries;
- }
-
- anv_batch_emit(batch, GENX(3DSTATE_URB_HS), urb) {
- urb.HSURBStartingAddress = push_constant_chunks;
- }
-
- anv_batch_emit(batch, GENX(3DSTATE_URB_DS), urb) {
- urb.DSURBStartingAddress = push_constant_chunks;
- }
-
- anv_batch_emit(batch, GENX(3DSTATE_URB_GS), urb) {
- urb.GSURBStartingAddress = push_constant_chunks + vs_chunks;
- urb.GSURBEntryAllocationSize = gs_size - 1;
- urb.GSNumberofURBEntries = nr_gs_entries;
+ for (int i = 0; i <= MESA_SHADER_GEOMETRY; i++) {
+ anv_batch_emit(batch, GENX(3DSTATE_URB_VS), urb) {
+ urb._3DCommandSubOpcode += i;
+ urb.VSURBStartingAddress = start[i];
+ urb.VSURBEntryAllocationSize = entry_size[i] - 1;
+ urb.VSNumberofURBEntries = entries[i];
+ }
}
}
-static inline void
+static void
emit_urb_setup(struct anv_pipeline *pipeline)
{
- unsigned vs_entry_size =
- (pipeline->active_stages & VK_SHADER_STAGE_VERTEX_BIT) ?
- get_vs_prog_data(pipeline)->base.urb_entry_size : 0;
- unsigned gs_entry_size =
- (pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT) ?
- get_gs_prog_data(pipeline)->base.urb_entry_size : 0;
+ unsigned entry_size[4];
+ for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
+ const struct brw_vue_prog_data *prog_data =
+ !anv_pipeline_has_stage(pipeline, i) ? NULL :
+ (const struct brw_vue_prog_data *) pipeline->shaders[i]->prog_data;
+
+ entry_size[i] = prog_data ? prog_data->urb_entry_size : 1;
+ }
genX(emit_urb_setup)(pipeline->device, &pipeline->batch,
- pipeline->active_stages, vs_entry_size, gs_entry_size,
- pipeline->urb.l3_config);
+ pipeline->urb.l3_config,
+ pipeline->active_stages, entry_size);
}
static void
emit_3dstate_sbe(struct anv_pipeline *pipeline)
{
- const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
- const struct brw_gs_prog_data *gs_prog_data = get_gs_prog_data(pipeline);
const struct brw_wm_prog_data *wm_prog_data = get_wm_prog_data(pipeline);
- const struct brw_vue_map *fs_input_map;
if (!anv_pipeline_has_stage(pipeline, MESA_SHADER_FRAGMENT)) {
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_SBE), sbe);
return;
}
- if (gs_prog_data)
- fs_input_map = &gs_prog_data->base.vue_map;
- else
- fs_input_map = &vs_prog_data->base.vue_map;
+ const struct brw_vue_map *fs_input_map =
+ &anv_pipeline_get_last_vue_prog_data(pipeline)->vue_map;
struct GENX(3DSTATE_SBE) sbe = {
GENX(3DSTATE_SBE_header),
# define swiz sbe
#endif
+ /* Skip the VUE header and position slots by default */
+ unsigned urb_entry_read_offset = 1;
int max_source_attr = 0;
for (int attr = 0; attr < VARYING_SLOT_MAX; attr++) {
int input_index = wm_prog_data->urb_setup[attr];
if (input_index < 0)
continue;
+ /* gl_Layer is stored in the VUE header */
+ if (attr == VARYING_SLOT_LAYER) {
+ urb_entry_read_offset = 0;
+ continue;
+ }
+
if (attr == VARYING_SLOT_PNTC) {
sbe.PointSpriteTextureCoordinateEnable = 1 << input_index;
continue;
swiz.Attribute[input_index].ComponentOverrideZ = true;
swiz.Attribute[input_index].ComponentOverrideW = true;
} else {
- assert(slot >= 2);
- const int source_attr = slot - 2;
- max_source_attr = MAX2(max_source_attr, source_attr);
/* We have to subtract two slots to accout for the URB entry output
* read offset in the VS and GS stages.
*/
+ const int source_attr = slot - 2 * urb_entry_read_offset;
+ assert(source_attr >= 0 && source_attr < 32);
+ max_source_attr = MAX2(max_source_attr, source_attr);
swiz.Attribute[input_index].SourceAttribute = source_attr;
}
}
- sbe.VertexURBEntryReadOffset = 1; /* Skip the VUE header and position slots */
+ sbe.VertexURBEntryReadOffset = urb_entry_read_offset;
sbe.VertexURBEntryReadLength = DIV_ROUND_UP(max_source_attr + 1, 2);
+#if GEN_GEN >= 8
+ sbe.ForceVertexURBEntryReadOffset = true;
+ sbe.ForceVertexURBEntryReadLength = true;
+#endif
uint32_t *dw = anv_batch_emit_dwords(&pipeline->batch,
GENX(3DSTATE_SBE_length));
+ if (!dw)
+ return;
GENX(3DSTATE_SBE_pack)(&pipeline->batch, dw, &sbe);
#if GEN_GEN >= 8
dw = anv_batch_emit_dwords(&pipeline->batch, GENX(3DSTATE_SBE_SWIZ_length));
+ if (!dw)
+ return;
GENX(3DSTATE_SBE_SWIZ_pack)(&pipeline->batch, dw, &swiz);
#endif
}
sf.TriangleStripListProvokingVertexSelect = 0;
sf.LineStripListProvokingVertexSelect = 0;
sf.TriangleFanProvokingVertexSelect = 1;
- sf.PointWidthSource = Vertex;
- sf.PointWidth = 1.0;
+
+ const struct brw_vue_prog_data *last_vue_prog_data =
+ anv_pipeline_get_last_vue_prog_data(pipeline);
+
+ if (last_vue_prog_data->vue_map.slots_valid & VARYING_BIT_PSIZ) {
+ sf.PointWidthSource = Vertex;
+ } else {
+ sf.PointWidthSource = State;
+ sf.PointWidth = 1.0;
+ }
#if GEN_GEN >= 8
struct GENX(3DSTATE_RASTER) raster = {
*/
#if GEN_GEN >= 8
raster.DXMultisampleRasterizationEnable = true;
+ /* NOTE: 3DSTATE_RASTER::ForcedSampleCount affects the BDW and SKL PMA fix
+ * computations. If we ever set this bit to a different value, they will
+ * need to be updated accordingly.
+ */
raster.ForcedSampleCount = FSC_NUMRASTSAMPLES_0;
raster.ForceMultisampling = false;
#else
/* Gen7 requires that we provide the depth format in 3DSTATE_SF so that it
* can get the depth offsets correct.
*/
- if (subpass->depth_stencil_attachment < pass->attachment_count) {
+ if (subpass->depth_stencil_attachment.attachment < pass->attachment_count) {
VkFormat vk_format =
- pass->attachments[subpass->depth_stencil_attachment].format;
+ pass->attachments[subpass->depth_stencil_attachment.attachment].format;
assert(vk_format_is_depth_or_stencil(vk_format));
if (vk_format_aspects(vk_format) & VK_IMAGE_ASPECT_DEPTH_BIT) {
enum isl_format isl_format =
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_MULTISAMPLE), ms) {
ms.NumberofMultisamples = log2_samples;
+ ms.PixelLocation = CENTER;
#if GEN_GEN >= 8
/* The PRM says that this bit is valid only for DX9:
*
* should not have any effect by setting or not setting this bit.
*/
ms.PixelPositionOffsetEnable = false;
- ms.PixelLocation = CENTER;
#else
- ms.PixelLocation = PIXLOC_CENTER;
switch (samples) {
case 1:
[VK_STENCIL_OP_DECREMENT_AND_WRAP] = STENCILOP_DECR,
};
+/* This function sanitizes the VkStencilOpState by looking at the compare ops
+ * and trying to determine whether or not a given stencil op can ever actually
+ * occur. Stencil ops which can never occur are set to VK_STENCIL_OP_KEEP.
+ * This function returns true if, after sanitation, any of the stencil ops are
+ * set to something other than VK_STENCIL_OP_KEEP.
+ */
+static bool
+sanitize_stencil_face(VkStencilOpState *face,
+ VkCompareOp depthCompareOp)
+{
+ /* If compareOp is ALWAYS then the stencil test will never fail and failOp
+ * will never happen. Set failOp to KEEP in this case.
+ */
+ if (face->compareOp == VK_COMPARE_OP_ALWAYS)
+ face->failOp = VK_STENCIL_OP_KEEP;
+
+ /* If compareOp is NEVER or depthCompareOp is NEVER then one of the depth
+ * or stencil tests will fail and passOp will never happen.
+ */
+ if (face->compareOp == VK_COMPARE_OP_NEVER ||
+ depthCompareOp == VK_COMPARE_OP_NEVER)
+ face->passOp = VK_STENCIL_OP_KEEP;
+
+ /* If compareOp is NEVER or depthCompareOp is ALWAYS then either the
+ * stencil test will fail or the depth test will pass. In either case,
+ * depthFailOp will never happen.
+ */
+ if (face->compareOp == VK_COMPARE_OP_NEVER ||
+ depthCompareOp == VK_COMPARE_OP_ALWAYS)
+ face->depthFailOp = VK_STENCIL_OP_KEEP;
+
+ return face->failOp != VK_STENCIL_OP_KEEP ||
+ face->depthFailOp != VK_STENCIL_OP_KEEP ||
+ face->passOp != VK_STENCIL_OP_KEEP;
+}
+
+/* Intel hardware is fairly sensitive to whether or not depth/stencil writes
+ * are enabled. In the presence of discards, it's fairly easy to get into the
+ * non-promoted case which means a fairly big performance hit. From the Iron
+ * Lake PRM, Vol 2, pt. 1, section 8.4.3.2, "Early Depth Test Cases":
+ *
+ * "Non-promoted depth (N) is active whenever the depth test can be done
+ * early but it cannot determine whether or not to write source depth to
+ * the depth buffer, therefore the depth write must be performed post pixel
+ * shader. This includes cases where the pixel shader can kill pixels,
+ * including via sampler chroma key, as well as cases where the alpha test
+ * function is enabled, which kills pixels based on a programmable alpha
+ * test. In this case, even if the depth test fails, the pixel cannot be
+ * killed if a stencil write is indicated. Whether or not the stencil write
+ * happens depends on whether or not the pixel is killed later. In these
+ * cases if stencil test fails and stencil writes are off, the pixels can
+ * also be killed early. If stencil writes are enabled, the pixels must be
+ * treated as Computed depth (described above)."
+ *
+ * The same thing as mentioned in the stencil case can happen in the depth
+ * case as well if it thinks it writes depth but, thanks to the depth test
+ * being GL_EQUAL, the write doesn't actually matter. A little extra work
+ * up-front to try and disable depth and stencil writes can make a big
+ * difference.
+ *
+ * Unfortunately, the way depth and stencil testing is specified, there are
+ * many case where, regardless of depth/stencil writes being enabled, nothing
+ * actually gets written due to some other bit of state being set. This
+ * function attempts to "sanitize" the depth stencil state and disable writes
+ * and sometimes even testing whenever possible.
+ */
+static void
+sanitize_ds_state(VkPipelineDepthStencilStateCreateInfo *state,
+ bool *stencilWriteEnable,
+ VkImageAspectFlags ds_aspects)
+{
+ *stencilWriteEnable = state->stencilTestEnable;
+
+ /* If the depth test is disabled, we won't be writing anything. */
+ if (!state->depthTestEnable)
+ state->depthWriteEnable = false;
+
+ /* The Vulkan spec requires that if either depth or stencil is not present,
+ * the pipeline is to act as if the test silently passes.
+ */
+ if (!(ds_aspects & VK_IMAGE_ASPECT_DEPTH_BIT)) {
+ state->depthWriteEnable = false;
+ state->depthCompareOp = VK_COMPARE_OP_ALWAYS;
+ }
+
+ if (!(ds_aspects & VK_IMAGE_ASPECT_STENCIL_BIT)) {
+ *stencilWriteEnable = false;
+ state->front.compareOp = VK_COMPARE_OP_ALWAYS;
+ state->back.compareOp = VK_COMPARE_OP_ALWAYS;
+ }
+
+ /* If the stencil test is enabled and always fails, then we will never get
+ * to the depth test so we can just disable the depth test entirely.
+ */
+ if (state->stencilTestEnable &&
+ state->front.compareOp == VK_COMPARE_OP_NEVER &&
+ state->back.compareOp == VK_COMPARE_OP_NEVER) {
+ state->depthTestEnable = false;
+ state->depthWriteEnable = false;
+ }
+
+ /* If depthCompareOp is EQUAL then the value we would be writing to the
+ * depth buffer is the same as the value that's already there so there's no
+ * point in writing it.
+ */
+ if (state->depthCompareOp == VK_COMPARE_OP_EQUAL)
+ state->depthWriteEnable = false;
+
+ /* If the stencil ops are such that we don't actually ever modify the
+ * stencil buffer, we should disable writes.
+ */
+ if (!sanitize_stencil_face(&state->front, state->depthCompareOp) &&
+ !sanitize_stencil_face(&state->back, state->depthCompareOp))
+ *stencilWriteEnable = false;
+
+ /* If the depth test always passes and we never write out depth, that's the
+ * same as if the depth test is disabled entirely.
+ */
+ if (state->depthCompareOp == VK_COMPARE_OP_ALWAYS &&
+ !state->depthWriteEnable)
+ state->depthTestEnable = false;
+
+ /* If the stencil test always passes and we never write out stencil, that's
+ * the same as if the stencil test is disabled entirely.
+ */
+ if (state->front.compareOp == VK_COMPARE_OP_ALWAYS &&
+ state->back.compareOp == VK_COMPARE_OP_ALWAYS &&
+ !*stencilWriteEnable)
+ state->stencilTestEnable = false;
+}
+
static void
emit_ds_state(struct anv_pipeline *pipeline,
- const VkPipelineDepthStencilStateCreateInfo *info,
+ const VkPipelineDepthStencilStateCreateInfo *pCreateInfo,
const struct anv_render_pass *pass,
const struct anv_subpass *subpass)
{
# define depth_stencil_dw pipeline->gen9.wm_depth_stencil
#endif
- if (info == NULL) {
+ if (pCreateInfo == NULL) {
/* We're going to OR this together with the dynamic state. We need
* to make sure it's initialized to something useful.
*/
+ pipeline->writes_stencil = false;
+ pipeline->stencil_test_enable = false;
+ pipeline->writes_depth = false;
+ pipeline->depth_test_enable = false;
memset(depth_stencil_dw, 0, sizeof(depth_stencil_dw));
return;
}
+ VkImageAspectFlags ds_aspects = 0;
+ if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
+ VkFormat depth_stencil_format =
+ pass->attachments[subpass->depth_stencil_attachment.attachment].format;
+ ds_aspects = vk_format_aspects(depth_stencil_format);
+ }
+
+ VkPipelineDepthStencilStateCreateInfo info = *pCreateInfo;
+ sanitize_ds_state(&info, &pipeline->writes_stencil, ds_aspects);
+ pipeline->stencil_test_enable = info.stencilTestEnable;
+ pipeline->writes_depth = info.depthWriteEnable;
+ pipeline->depth_test_enable = info.depthTestEnable;
+
/* VkBool32 depthBoundsTestEnable; // optional (depth_bounds_test) */
#if GEN_GEN <= 7
#else
struct GENX(3DSTATE_WM_DEPTH_STENCIL) depth_stencil = {
#endif
- .DepthTestEnable = info->depthTestEnable,
- .DepthBufferWriteEnable = info->depthWriteEnable,
- .DepthTestFunction = vk_to_gen_compare_op[info->depthCompareOp],
+ .DepthTestEnable = info.depthTestEnable,
+ .DepthBufferWriteEnable = info.depthWriteEnable,
+ .DepthTestFunction = vk_to_gen_compare_op[info.depthCompareOp],
.DoubleSidedStencilEnable = true,
- .StencilTestEnable = info->stencilTestEnable,
- .StencilBufferWriteEnable = info->stencilTestEnable,
- .StencilFailOp = vk_to_gen_stencil_op[info->front.failOp],
- .StencilPassDepthPassOp = vk_to_gen_stencil_op[info->front.passOp],
- .StencilPassDepthFailOp = vk_to_gen_stencil_op[info->front.depthFailOp],
- .StencilTestFunction = vk_to_gen_compare_op[info->front.compareOp],
- .BackfaceStencilFailOp = vk_to_gen_stencil_op[info->back.failOp],
- .BackfaceStencilPassDepthPassOp = vk_to_gen_stencil_op[info->back.passOp],
- .BackfaceStencilPassDepthFailOp =vk_to_gen_stencil_op[info->back.depthFailOp],
- .BackfaceStencilTestFunction = vk_to_gen_compare_op[info->back.compareOp],
+ .StencilTestEnable = info.stencilTestEnable,
+ .StencilFailOp = vk_to_gen_stencil_op[info.front.failOp],
+ .StencilPassDepthPassOp = vk_to_gen_stencil_op[info.front.passOp],
+ .StencilPassDepthFailOp = vk_to_gen_stencil_op[info.front.depthFailOp],
+ .StencilTestFunction = vk_to_gen_compare_op[info.front.compareOp],
+ .BackfaceStencilFailOp = vk_to_gen_stencil_op[info.back.failOp],
+ .BackfaceStencilPassDepthPassOp = vk_to_gen_stencil_op[info.back.passOp],
+ .BackfaceStencilPassDepthFailOp =vk_to_gen_stencil_op[info.back.depthFailOp],
+ .BackfaceStencilTestFunction = vk_to_gen_compare_op[info.back.compareOp],
};
- VkImageAspectFlags aspects = 0;
- if (subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED) {
- VkFormat depth_stencil_format =
- pass->attachments[subpass->depth_stencil_attachment].format;
- aspects = vk_format_aspects(depth_stencil_format);
- }
-
- /* The Vulkan spec requires that if either depth or stencil is not present,
- * the pipeline is to act as if the test silently passes.
- */
- if (!(aspects & VK_IMAGE_ASPECT_DEPTH_BIT)) {
- depth_stencil.DepthBufferWriteEnable = false;
- depth_stencil.DepthTestFunction = PREFILTEROPALWAYS;
- }
-
- if (!(aspects & VK_IMAGE_ASPECT_STENCIL_BIT)) {
- depth_stencil.StencilBufferWriteEnable = false;
- depth_stencil.StencilTestFunction = PREFILTEROPALWAYS;
- depth_stencil.BackfaceStencilTestFunction = PREFILTEROPALWAYS;
- }
-
- /* From the Broadwell PRM:
- *
- * "If Depth_Test_Enable = 1 AND Depth_Test_func = EQUAL, the
- * Depth_Write_Enable must be set to 0."
- */
- if (info->depthTestEnable && info->depthCompareOp == VK_COMPARE_OP_EQUAL)
- depth_stencil.DepthBufferWriteEnable = false;
-
#if GEN_GEN <= 7
GENX(DEPTH_STENCIL_STATE_pack)(NULL, depth_stencil_dw, &depth_stencil);
#else
{
struct anv_device *device = pipeline->device;
- const uint32_t num_dwords = GENX(BLEND_STATE_length);
- pipeline->blend_state =
- anv_state_pool_alloc(&device->dynamic_state_pool, num_dwords * 4, 64);
struct GENX(BLEND_STATE) blend_state = {
#if GEN_GEN >= 8
.AlphaToCoverageEnable = ms_info && ms_info->alphaToCoverageEnable,
.AlphaToOneEnable = ms_info && ms_info->alphaToOneEnable,
-#else
- /* Make sure it gets zeroed */
- .Entry = { { 0, }, },
#endif
};
- /* Default everything to disabled */
- for (uint32_t i = 0; i < 8; i++) {
- blend_state.Entry[i].WriteDisableAlpha = true;
- blend_state.Entry[i].WriteDisableRed = true;
- blend_state.Entry[i].WriteDisableGreen = true;
- blend_state.Entry[i].WriteDisableBlue = true;
- }
-
uint32_t surface_count = 0;
struct anv_pipeline_bind_map *map;
if (anv_pipeline_has_stage(pipeline, MESA_SHADER_FRAGMENT)) {
surface_count = map->surface_count;
}
+ const uint32_t num_dwords = GENX(BLEND_STATE_length) +
+ GENX(BLEND_STATE_ENTRY_length) * surface_count;
+ pipeline->blend_state =
+ anv_state_pool_alloc(&device->dynamic_state_pool, num_dwords * 4, 64);
+
bool has_writeable_rt = false;
+ uint32_t *state_pos = pipeline->blend_state.map;
+ state_pos += GENX(BLEND_STATE_length);
+#if GEN_GEN >= 8
+ struct GENX(BLEND_STATE_ENTRY) bs0 = { 0 };
+#endif
for (unsigned i = 0; i < surface_count; i++) {
struct anv_pipeline_binding *binding = &map->surface_to_descriptor[i];
/* We can have at most 8 attachments */
assert(i < 8);
- if (binding->index >= info->attachmentCount)
+ if (info == NULL || binding->index >= info->attachmentCount) {
+ /* Default everything to disabled */
+ struct GENX(BLEND_STATE_ENTRY) entry = {
+ .WriteDisableAlpha = true,
+ .WriteDisableRed = true,
+ .WriteDisableGreen = true,
+ .WriteDisableBlue = true,
+ };
+ GENX(BLEND_STATE_ENTRY_pack)(NULL, state_pos, &entry);
+ state_pos += GENX(BLEND_STATE_ENTRY_length);
continue;
+ }
assert(binding->binding == 0);
const VkPipelineColorBlendAttachmentState *a =
&info->pAttachments[binding->index];
- blend_state.Entry[i] = (struct GENX(BLEND_STATE_ENTRY)) {
+ struct GENX(BLEND_STATE_ENTRY) entry = {
#if GEN_GEN < 8
.AlphaToCoverageEnable = ms_info && ms_info->alphaToCoverageEnable,
.AlphaToOneEnable = ms_info && ms_info->alphaToOneEnable,
#if GEN_GEN >= 8
blend_state.IndependentAlphaBlendEnable = true;
#else
- blend_state.Entry[i].IndependentAlphaBlendEnable = true;
+ entry.IndependentAlphaBlendEnable = true;
#endif
}
*/
if (a->colorBlendOp == VK_BLEND_OP_MIN ||
a->colorBlendOp == VK_BLEND_OP_MAX) {
- blend_state.Entry[i].SourceBlendFactor = BLENDFACTOR_ONE;
- blend_state.Entry[i].DestinationBlendFactor = BLENDFACTOR_ONE;
+ entry.SourceBlendFactor = BLENDFACTOR_ONE;
+ entry.DestinationBlendFactor = BLENDFACTOR_ONE;
}
if (a->alphaBlendOp == VK_BLEND_OP_MIN ||
a->alphaBlendOp == VK_BLEND_OP_MAX) {
- blend_state.Entry[i].SourceAlphaBlendFactor = BLENDFACTOR_ONE;
- blend_state.Entry[i].DestinationAlphaBlendFactor = BLENDFACTOR_ONE;
+ entry.SourceAlphaBlendFactor = BLENDFACTOR_ONE;
+ entry.DestinationAlphaBlendFactor = BLENDFACTOR_ONE;
}
+ GENX(BLEND_STATE_ENTRY_pack)(NULL, state_pos, &entry);
+ state_pos += GENX(BLEND_STATE_ENTRY_length);
+#if GEN_GEN >= 8
+ if (i == 0)
+ bs0 = entry;
+#endif
}
#if GEN_GEN >= 8
- struct GENX(BLEND_STATE_ENTRY) *bs0 = &blend_state.Entry[0];
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_PS_BLEND), blend) {
blend.AlphaToCoverageEnable = blend_state.AlphaToCoverageEnable;
blend.HasWriteableRT = has_writeable_rt;
- blend.ColorBufferBlendEnable = bs0->ColorBufferBlendEnable;
- blend.SourceAlphaBlendFactor = bs0->SourceAlphaBlendFactor;
- blend.DestinationAlphaBlendFactor = bs0->DestinationAlphaBlendFactor;
- blend.SourceBlendFactor = bs0->SourceBlendFactor;
- blend.DestinationBlendFactor = bs0->DestinationBlendFactor;
+ blend.ColorBufferBlendEnable = bs0.ColorBufferBlendEnable;
+ blend.SourceAlphaBlendFactor = bs0.SourceAlphaBlendFactor;
+ blend.DestinationAlphaBlendFactor = bs0.DestinationAlphaBlendFactor;
+ blend.SourceBlendFactor = bs0.SourceBlendFactor;
+ blend.DestinationBlendFactor = bs0.DestinationBlendFactor;
blend.AlphaTestEnable = false;
blend.IndependentAlphaBlendEnable =
blend_state.IndependentAlphaBlendEnable;
#endif
GENX(BLEND_STATE_pack)(NULL, pipeline->blend_state.map, &blend_state);
- if (!device->info.has_llc)
- anv_state_clflush(pipeline->blend_state);
+ anv_state_flush(device, pipeline->blend_state);
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_BLEND_STATE_POINTERS), bsp) {
bsp.BlendStatePointer = pipeline->blend_state.offset;
(void) wm_prog_data;
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_CLIP), clip) {
clip.ClipEnable = true;
+ clip.StatisticsEnable = true;
clip.EarlyCullEnable = true;
clip.APIMode = APIMODE_D3D,
clip.ViewportXYClipTestEnable = true;
clip.MinimumPointWidth = 0.125;
clip.MaximumPointWidth = 255.875;
- clip.MaximumVPIndex = (vp_info ? vp_info->viewportCount : 1) - 1;
+
+ const struct brw_vue_prog_data *last =
+ anv_pipeline_get_last_vue_prog_data(pipeline);
+
+ /* From the Vulkan 1.0.45 spec:
+ *
+ * "If the last active vertex processing stage shader entry point's
+ * interface does not include a variable decorated with
+ * ViewportIndex, then the first viewport is used."
+ */
+ if (vp_info && (last->vue_map.slots_valid & VARYING_BIT_VIEWPORT)) {
+ clip.MaximumVPIndex = vp_info->viewportCount - 1;
+ } else {
+ clip.MaximumVPIndex = 0;
+ }
+
+ /* From the Vulkan 1.0.45 spec:
+ *
+ * "If the last active vertex processing stage shader entry point's
+ * interface does not include a variable decorated with Layer, then
+ * the first layer is used."
+ */
+ clip.ForceZeroRTAIndexEnable =
+ !(last->vue_map.slots_valid & VARYING_BIT_LAYER);
#if GEN_GEN == 7
clip.FrontWinding = vk_to_gen_front_face[rs_info->frontFace];
clip.CullMode = vk_to_gen_cullmode[rs_info->cullMode];
clip.ViewportZClipTestEnable = !pipeline->depth_clamp_enable;
+ if (last) {
+ clip.UserClipDistanceClipTestEnableBitmask = last->clip_distance_mask;
+ clip.UserClipDistanceCullTestEnableBitmask = last->cull_distance_mask;
+ }
#else
clip.NonPerspectiveBarycentricEnable = wm_prog_data ?
- (wm_prog_data->barycentric_interp_modes & 0x38) != 0 : 0;
+ (wm_prog_data->barycentric_interp_modes &
+ BRW_BARYCENTRIC_NONPERSPECTIVE_BITS) != 0 : 0;
#endif
}
}
}
}
-static inline uint32_t
+static uint32_t
get_sampler_count(const struct anv_shader_bin *bin)
{
- return DIV_ROUND_UP(bin->bind_map.sampler_count, 4);
+ uint32_t count_by_4 = DIV_ROUND_UP(bin->bind_map.sampler_count, 4);
+
+ /* We can potentially have way more than 32 samplers and that's ok.
+ * However, the 3DSTATE_XS packets only have 3 bits to specify how
+ * many to pre-fetch and all values above 4 are marked reserved.
+ */
+ return MIN2(count_by_4, 4);
}
-static inline uint32_t
+static uint32_t
get_binding_table_entry_count(const struct anv_shader_bin *bin)
{
return DIV_ROUND_UP(bin->bind_map.surface_count, 32);
}
-static inline struct anv_address
+static struct anv_address
get_scratch_address(struct anv_pipeline *pipeline,
gl_shader_stage stage,
const struct anv_shader_bin *bin)
};
}
-static inline uint32_t
+static uint32_t
get_scratch_space(const struct anv_shader_bin *bin)
{
return ffs(bin->prog_data->total_scratch / 2048);
}
-static inline uint32_t
-get_urb_output_offset()
-{
- /* Skip the VUE header and position slots */
- return 1;
-}
-
-static inline uint32_t
-get_urb_output_length(const struct anv_shader_bin *bin)
-{
- const struct brw_vue_prog_data *prog_data =
- (const struct brw_vue_prog_data *)bin->prog_data;
-
- return (prog_data->vue_map.num_slots + 1) / 2 - get_urb_output_offset();
-}
-
static void
emit_3dstate_vs(struct anv_pipeline *pipeline)
{
assert(anv_pipeline_has_stage(pipeline, MESA_SHADER_VERTEX));
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VS), vs) {
- vs.FunctionEnable = true;
+ vs.Enable = true;
vs.StatisticsEnable = true;
vs.KernelStartPointer = vs_bin->kernel.offset;
#if GEN_GEN >= 8
#endif
assert(!vs_prog_data->base.base.use_alt_mode);
+#if GEN_GEN < 11
vs.SingleVertexDispatch = false;
+#endif
vs.VectorMaskEnable = false;
vs.SamplerCount = get_sampler_count(vs_bin);
vs.BindingTableEntryCount = get_binding_table_entry_count(vs_bin);
vs_prog_data->base.base.dispatch_grf_start_reg;
#if GEN_GEN >= 8
- vs.VertexURBEntryOutputReadOffset = get_urb_output_offset();
- vs.VertexURBEntryOutputLength = get_urb_output_length(vs_bin);
-
- /* TODO */
- vs.UserClipDistanceClipTestEnableBitmask = 0;
- vs.UserClipDistanceCullTestEnableBitmask = 0;
+ vs.UserClipDistanceClipTestEnableBitmask =
+ vs_prog_data->base.clip_distance_mask;
+ vs.UserClipDistanceCullTestEnableBitmask =
+ vs_prog_data->base.cull_distance_mask;
#endif
vs.PerThreadScratchSpace = get_scratch_space(vs_bin);
}
}
+static void
+emit_3dstate_hs_te_ds(struct anv_pipeline *pipeline,
+ const VkPipelineTessellationStateCreateInfo *tess_info)
+{
+ if (!anv_pipeline_has_stage(pipeline, MESA_SHADER_TESS_EVAL)) {
+ anv_batch_emit(&pipeline->batch, GENX(3DSTATE_HS), hs);
+ anv_batch_emit(&pipeline->batch, GENX(3DSTATE_TE), te);
+ anv_batch_emit(&pipeline->batch, GENX(3DSTATE_DS), ds);
+ return;
+ }
+
+ const struct gen_device_info *devinfo = &pipeline->device->info;
+ const struct anv_shader_bin *tcs_bin =
+ pipeline->shaders[MESA_SHADER_TESS_CTRL];
+ const struct anv_shader_bin *tes_bin =
+ pipeline->shaders[MESA_SHADER_TESS_EVAL];
+
+ const struct brw_tcs_prog_data *tcs_prog_data = get_tcs_prog_data(pipeline);
+ const struct brw_tes_prog_data *tes_prog_data = get_tes_prog_data(pipeline);
+
+ anv_batch_emit(&pipeline->batch, GENX(3DSTATE_HS), hs) {
+ hs.Enable = true;
+ hs.StatisticsEnable = true;
+ hs.KernelStartPointer = tcs_bin->kernel.offset;
+
+ hs.SamplerCount = get_sampler_count(tcs_bin);
+ hs.BindingTableEntryCount = get_binding_table_entry_count(tcs_bin);
+ hs.MaximumNumberofThreads = devinfo->max_tcs_threads - 1;
+ hs.IncludeVertexHandles = true;
+ hs.InstanceCount = tcs_prog_data->instances - 1;
+
+ hs.VertexURBEntryReadLength = 0;
+ hs.VertexURBEntryReadOffset = 0;
+ hs.DispatchGRFStartRegisterForURBData =
+ tcs_prog_data->base.base.dispatch_grf_start_reg;
+
+ hs.PerThreadScratchSpace = get_scratch_space(tcs_bin);
+ hs.ScratchSpaceBasePointer =
+ get_scratch_address(pipeline, MESA_SHADER_TESS_CTRL, tcs_bin);
+ }
+
+ const VkPipelineTessellationDomainOriginStateCreateInfoKHR *domain_origin_state =
+ tess_info ? vk_find_struct_const(tess_info, PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO_KHR) : NULL;
+
+ VkTessellationDomainOriginKHR uv_origin =
+ domain_origin_state ? domain_origin_state->domainOrigin :
+ VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT_KHR;
+
+ anv_batch_emit(&pipeline->batch, GENX(3DSTATE_TE), te) {
+ te.Partitioning = tes_prog_data->partitioning;
+
+ if (uv_origin == VK_TESSELLATION_DOMAIN_ORIGIN_LOWER_LEFT_KHR) {
+ te.OutputTopology = tes_prog_data->output_topology;
+ } else {
+ /* When the origin is upper-left, we have to flip the winding order */
+ if (tes_prog_data->output_topology == OUTPUT_TRI_CCW) {
+ te.OutputTopology = OUTPUT_TRI_CW;
+ } else if (tes_prog_data->output_topology == OUTPUT_TRI_CW) {
+ te.OutputTopology = OUTPUT_TRI_CCW;
+ } else {
+ te.OutputTopology = tes_prog_data->output_topology;
+ }
+ }
+
+ te.TEDomain = tes_prog_data->domain;
+ te.TEEnable = true;
+ te.MaximumTessellationFactorOdd = 63.0;
+ te.MaximumTessellationFactorNotOdd = 64.0;
+ }
+
+ anv_batch_emit(&pipeline->batch, GENX(3DSTATE_DS), ds) {
+ ds.Enable = true;
+ ds.StatisticsEnable = true;
+ ds.KernelStartPointer = tes_bin->kernel.offset;
+
+ ds.SamplerCount = get_sampler_count(tes_bin);
+ ds.BindingTableEntryCount = get_binding_table_entry_count(tes_bin);
+ ds.MaximumNumberofThreads = devinfo->max_tes_threads - 1;
+
+ ds.ComputeWCoordinateEnable =
+ tes_prog_data->domain == BRW_TESS_DOMAIN_TRI;
+
+ ds.PatchURBEntryReadLength = tes_prog_data->base.urb_read_length;
+ ds.PatchURBEntryReadOffset = 0;
+ ds.DispatchGRFStartRegisterForURBData =
+ tes_prog_data->base.base.dispatch_grf_start_reg;
+
+#if GEN_GEN >= 8
+#if GEN_GEN < 11
+ ds.DispatchMode =
+ tes_prog_data->base.dispatch_mode == DISPATCH_MODE_SIMD8 ?
+ DISPATCH_MODE_SIMD8_SINGLE_PATCH :
+ DISPATCH_MODE_SIMD4X2;
+#else
+ assert(tes_prog_data->base.dispatch_mode == DISPATCH_MODE_SIMD8);
+ ds.DispatchMode = DISPATCH_MODE_SIMD8_SINGLE_PATCH;
+#endif
+
+ ds.UserClipDistanceClipTestEnableBitmask =
+ tes_prog_data->base.clip_distance_mask;
+ ds.UserClipDistanceCullTestEnableBitmask =
+ tes_prog_data->base.cull_distance_mask;
+#endif
+
+ ds.PerThreadScratchSpace = get_scratch_space(tes_bin);
+ ds.ScratchSpaceBasePointer =
+ get_scratch_address(pipeline, MESA_SHADER_TESS_EVAL, tes_bin);
+ }
+}
+
static void
emit_3dstate_gs(struct anv_pipeline *pipeline)
{
const struct brw_gs_prog_data *gs_prog_data = get_gs_prog_data(pipeline);
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_GS), gs) {
- gs.FunctionEnable = true;
+ gs.Enable = true;
gs.StatisticsEnable = true;
gs.KernelStartPointer = gs_bin->kernel.offset;
gs.DispatchMode = gs_prog_data->base.dispatch_mode;
gs.ControlDataFormat = gs_prog_data->control_data_format;
gs.ControlDataHeaderSize = gs_prog_data->control_data_header_size_hwords;
gs.InstanceControl = MAX2(gs_prog_data->invocations, 1) - 1;
-#if GEN_GEN >= 8 || GEN_IS_HASWELL
gs.ReorderMode = TRAILING;
-#else
- gs.ReorderEnable = true;
-#endif
#if GEN_GEN >= 8
gs.ExpectedVertexCount = gs_prog_data->vertices_in;
gs_prog_data->base.base.dispatch_grf_start_reg;
#if GEN_GEN >= 8
- gs.VertexURBEntryOutputReadOffset = get_urb_output_offset();
- gs.VertexURBEntryOutputLength = get_urb_output_length(gs_bin);
-
- /* TODO */
- gs.UserClipDistanceClipTestEnableBitmask = 0;
- gs.UserClipDistanceCullTestEnableBitmask = 0;
+ gs.UserClipDistanceClipTestEnableBitmask =
+ gs_prog_data->base.clip_distance_mask;
+ gs.UserClipDistanceCullTestEnableBitmask =
+ gs_prog_data->base.cull_distance_mask;
#endif
gs.PerThreadScratchSpace = get_scratch_space(gs_bin);
}
}
+static bool
+has_color_buffer_write_enabled(const struct anv_pipeline *pipeline,
+ const VkPipelineColorBlendStateCreateInfo *blend)
+{
+ const struct anv_shader_bin *shader_bin =
+ pipeline->shaders[MESA_SHADER_FRAGMENT];
+ if (!shader_bin)
+ return false;
+
+ const struct anv_pipeline_bind_map *bind_map = &shader_bin->bind_map;
+ for (int i = 0; i < bind_map->surface_count; i++) {
+ struct anv_pipeline_binding *binding = &bind_map->surface_to_descriptor[i];
+
+ if (binding->set != ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS)
+ continue;
+
+ if (binding->index == UINT32_MAX)
+ continue;
+
+ if (blend && blend->pAttachments[binding->index].colorWriteMask != 0)
+ return true;
+ }
+
+ return false;
+}
+
static void
-emit_3dstate_wm(struct anv_pipeline *pipeline,
+emit_3dstate_wm(struct anv_pipeline *pipeline, struct anv_subpass *subpass,
+ const VkPipelineColorBlendStateCreateInfo *blend,
const VkPipelineMultisampleStateCreateInfo *multisample)
{
const struct brw_wm_prog_data *wm_prog_data = get_wm_prog_data(pipeline);
wm_prog_data->barycentric_interp_modes;
#if GEN_GEN < 8
- /* FIXME: This needs a lot more work, cf gen7 upload_wm_state(). */
- wm.ThreadDispatchEnable = true;
-
- wm.PixelShaderKillsPixel = wm_prog_data->uses_kill;
wm.PixelShaderComputedDepthMode = wm_prog_data->computed_depth_mode;
wm.PixelShaderUsesSourceDepth = wm_prog_data->uses_src_depth;
wm.PixelShaderUsesSourceW = wm_prog_data->uses_src_w;
wm.PixelShaderUsesInputCoverageMask = wm_prog_data->uses_sample_mask;
+ /* If the subpass has a depth or stencil self-dependency, then we
+ * need to force the hardware to do the depth/stencil write *after*
+ * fragment shader execution. Otherwise, the writes may hit memory
+ * before we get around to fetching from the input attachment and we
+ * may get the depth or stencil value from the current draw rather
+ * than the previous one.
+ */
+ wm.PixelShaderKillsPixel = subpass->has_ds_self_dep ||
+ wm_prog_data->uses_kill;
+
+ if (wm.PixelShaderComputedDepthMode != PSCDEPTH_OFF ||
+ wm_prog_data->has_side_effects ||
+ wm.PixelShaderKillsPixel ||
+ has_color_buffer_write_enabled(pipeline, blend))
+ wm.ThreadDispatchEnable = true;
+
if (samples > 1) {
wm.MultisampleRasterizationMode = MSRASTMODE_ON_PATTERN;
if (wm_prog_data->persample_dispatch) {
}
}
+UNUSED static bool
+is_dual_src_blend_factor(VkBlendFactor factor)
+{
+ return factor == VK_BLEND_FACTOR_SRC1_COLOR ||
+ factor == VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR ||
+ factor == VK_BLEND_FACTOR_SRC1_ALPHA ||
+ factor == VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA;
+}
+
static void
-emit_3dstate_ps(struct anv_pipeline *pipeline)
+emit_3dstate_ps(struct anv_pipeline *pipeline,
+ const VkPipelineColorBlendStateCreateInfo *blend)
{
MAYBE_UNUSED const struct gen_device_info *devinfo = &pipeline->device->info;
const struct anv_shader_bin *fs_bin =
const struct brw_wm_prog_data *wm_prog_data = get_wm_prog_data(pipeline);
+#if GEN_GEN < 8
+ /* The hardware wedges if you have this bit set but don't turn on any dual
+ * source blend factors.
+ */
+ bool dual_src_blend = false;
+ if (wm_prog_data->dual_src_blend && blend) {
+ for (uint32_t i = 0; i < blend->attachmentCount; i++) {
+ const VkPipelineColorBlendAttachmentState *bstate =
+ &blend->pAttachments[i];
+
+ if (bstate->blendEnable &&
+ (is_dual_src_blend_factor(bstate->srcColorBlendFactor) ||
+ is_dual_src_blend_factor(bstate->dstColorBlendFactor) ||
+ is_dual_src_blend_factor(bstate->srcAlphaBlendFactor) ||
+ is_dual_src_blend_factor(bstate->dstAlphaBlendFactor))) {
+ dual_src_blend = true;
+ break;
+ }
+ }
+ }
+#endif
+
anv_batch_emit(&pipeline->batch, GENX(3DSTATE_PS), ps) {
- ps.KernelStartPointer0 = fs_bin->kernel.offset;
- ps.KernelStartPointer1 = 0;
- ps.KernelStartPointer2 = fs_bin->kernel.offset +
- wm_prog_data->prog_offset_2;
ps._8PixelDispatchEnable = wm_prog_data->dispatch_8;
ps._16PixelDispatchEnable = wm_prog_data->dispatch_16;
- ps._32PixelDispatchEnable = false;
+ ps._32PixelDispatchEnable = wm_prog_data->dispatch_32;
+
+ ps.KernelStartPointer0 = fs_bin->kernel.offset +
+ brw_wm_prog_data_prog_offset(wm_prog_data, ps, 0);
+ ps.KernelStartPointer1 = fs_bin->kernel.offset +
+ brw_wm_prog_data_prog_offset(wm_prog_data, ps, 1);
+ ps.KernelStartPointer2 = fs_bin->kernel.offset +
+ brw_wm_prog_data_prog_offset(wm_prog_data, ps, 2);
ps.SingleProgramFlow = false;
ps.VectorMaskEnable = true;
ps.SamplerCount = get_sampler_count(fs_bin);
ps.BindingTableEntryCount = get_binding_table_entry_count(fs_bin);
- ps.PushConstantEnable = wm_prog_data->base.nr_params > 0;
+ ps.PushConstantEnable = wm_prog_data->base.nr_params > 0 ||
+ wm_prog_data->base.ubo_ranges[0].length;
ps.PositionXYOffsetSelect = wm_prog_data->uses_pos_offset ?
POSOFFSET_SAMPLE: POSOFFSET_NONE;
#if GEN_GEN < 8
ps.AttributeEnable = wm_prog_data->num_varying_inputs > 0;
ps.oMaskPresenttoRenderTarget = wm_prog_data->uses_omask;
- ps.DualSourceBlendEnable = wm_prog_data->dual_src_blend;
+ ps.DualSourceBlendEnable = dual_src_blend;
#endif
#if GEN_IS_HASWELL
#endif
ps.DispatchGRFStartRegisterForConstantSetupData0 =
- wm_prog_data->base.dispatch_grf_start_reg;
- ps.DispatchGRFStartRegisterForConstantSetupData1 = 0;
+ brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data, ps, 0);
+ ps.DispatchGRFStartRegisterForConstantSetupData1 =
+ brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data, ps, 1);
ps.DispatchGRFStartRegisterForConstantSetupData2 =
- wm_prog_data->dispatch_grf_start_reg_2;
+ brw_wm_prog_data_dispatch_grf_start_reg(wm_prog_data, ps, 2);
ps.PerThreadScratchSpace = get_scratch_space(fs_bin);
ps.ScratchSpaceBasePointer =
#if GEN_GEN >= 8
static void
-emit_3dstate_ps_extra(struct anv_pipeline *pipeline)
+emit_3dstate_ps_extra(struct anv_pipeline *pipeline,
+ struct anv_subpass *subpass,
+ const VkPipelineColorBlendStateCreateInfo *blend)
{
const struct brw_wm_prog_data *wm_prog_data = get_wm_prog_data(pipeline);
ps.AttributeEnable = wm_prog_data->num_varying_inputs > 0;
ps.oMaskPresenttoRenderTarget = wm_prog_data->uses_omask;
ps.PixelShaderIsPerSample = wm_prog_data->persample_dispatch;
- ps.PixelShaderKillsPixel = wm_prog_data->uses_kill;
ps.PixelShaderComputedDepthMode = wm_prog_data->computed_depth_mode;
ps.PixelShaderUsesSourceDepth = wm_prog_data->uses_src_depth;
ps.PixelShaderUsesSourceW = wm_prog_data->uses_src_w;
+ /* If the subpass has a depth or stencil self-dependency, then we need
+ * to force the hardware to do the depth/stencil write *after* fragment
+ * shader execution. Otherwise, the writes may hit memory before we get
+ * around to fetching from the input attachment and we may get the depth
+ * or stencil value from the current draw rather than the previous one.
+ */
+ ps.PixelShaderKillsPixel = subpass->has_ds_self_dep ||
+ wm_prog_data->uses_kill;
+
+ /* The stricter cross-primitive coherency guarantees that the hardware
+ * gives us with the "Accesses UAV" bit set for at least one shader stage
+ * and the "UAV coherency required" bit set on the 3DPRIMITIVE command are
+ * redundant within the current image, atomic counter and SSBO GL APIs,
+ * which all have very loose ordering and coherency requirements and
+ * generally rely on the application to insert explicit barriers when a
+ * shader invocation is expected to see the memory writes performed by the
+ * invocations of some previous primitive. Regardless of the value of
+ * "UAV coherency required", the "Accesses UAV" bits will implicitly cause
+ * an in most cases useless DC flush when the lowermost stage with the bit
+ * set finishes execution.
+ *
+ * It would be nice to disable it, but in some cases we can't because on
+ * Gen8+ it also has an influence on rasterization via the PS UAV-only
+ * signal (which could be set independently from the coherency mechanism
+ * in the 3DSTATE_WM command on Gen7), and because in some cases it will
+ * determine whether the hardware skips execution of the fragment shader
+ * or not via the ThreadDispatchEnable signal. However if we know that
+ * GEN8_PS_BLEND_HAS_WRITEABLE_RT is going to be set and
+ * GEN8_PSX_PIXEL_SHADER_NO_RT_WRITE is not set it shouldn't make any
+ * difference so we may just disable it here.
+ *
+ * Gen8 hardware tries to compute ThreadDispatchEnable for us but doesn't
+ * take into account KillPixels when no depth or stencil writes are
+ * enabled. In order for occlusion queries to work correctly with no
+ * attachments, we need to force-enable here.
+ */
+ if ((wm_prog_data->has_side_effects || wm_prog_data->uses_kill) &&
+ !has_color_buffer_write_enabled(pipeline, blend))
+ ps.PixelShaderHasUAV = true;
+
#if GEN_GEN >= 9
+ ps.PixelShaderComputesStencil = wm_prog_data->computed_stencil;
ps.PixelShaderPullsBary = wm_prog_data->pulls_bary;
ps.InputCoverageMaskState = wm_prog_data->uses_sample_mask ?
ICMS_INNER_CONSERVATIVE : ICMS_NONE;
}
#endif
+static void
+emit_3dstate_vf_statistics(struct anv_pipeline *pipeline)
+{
+ anv_batch_emit(&pipeline->batch, GENX(3DSTATE_VF_STATISTICS), vfs) {
+ vfs.StatisticsEnable = true;
+ }
+}
+
+static void
+compute_kill_pixel(struct anv_pipeline *pipeline,
+ const VkPipelineMultisampleStateCreateInfo *ms_info,
+ const struct anv_subpass *subpass)
+{
+ if (!anv_pipeline_has_stage(pipeline, MESA_SHADER_FRAGMENT)) {
+ pipeline->kill_pixel = false;
+ return;
+ }
+
+ const struct brw_wm_prog_data *wm_prog_data = get_wm_prog_data(pipeline);
+
+ /* This computes the KillPixel portion of the computation for whether or
+ * not we want to enable the PMA fix on gen8 or gen9. It's given by this
+ * chunk of the giant formula:
+ *
+ * (3DSTATE_PS_EXTRA::PixelShaderKillsPixels ||
+ * 3DSTATE_PS_EXTRA::oMask Present to RenderTarget ||
+ * 3DSTATE_PS_BLEND::AlphaToCoverageEnable ||
+ * 3DSTATE_PS_BLEND::AlphaTestEnable ||
+ * 3DSTATE_WM_CHROMAKEY::ChromaKeyKillEnable)
+ *
+ * 3DSTATE_WM_CHROMAKEY::ChromaKeyKillEnable is always false and so is
+ * 3DSTATE_PS_BLEND::AlphaTestEnable since Vulkan doesn't have a concept
+ * of an alpha test.
+ */
+ pipeline->kill_pixel =
+ subpass->has_ds_self_dep || wm_prog_data->uses_kill ||
+ wm_prog_data->uses_omask ||
+ (ms_info && ms_info->alphaToCoverageEnable);
+}
+
static VkResult
genX(graphics_pipeline_create)(
VkDevice _device,
emit_ds_state(pipeline, pCreateInfo->pDepthStencilState, pass, subpass);
emit_cb_state(pipeline, pCreateInfo->pColorBlendState,
pCreateInfo->pMultisampleState);
+ compute_kill_pixel(pipeline, pCreateInfo->pMultisampleState, subpass);
emit_urb_setup(pipeline);
* whole fixed function pipeline" means to emit a PIPE_CONTROL with the "CS
* Stall" bit set.
*/
- if (!brw->is_haswell && !brw->is_baytrail)
+ if (!device->info.is_haswell && !device->info.is_baytrail)
gen7_emit_vs_workaround_flush(brw);
#endif
emit_3dstate_vs(pipeline);
+ emit_3dstate_hs_te_ds(pipeline, pCreateInfo->pTessellationState);
emit_3dstate_gs(pipeline);
emit_3dstate_sbe(pipeline);
- emit_3dstate_wm(pipeline, pCreateInfo->pMultisampleState);
- emit_3dstate_ps(pipeline);
+ emit_3dstate_wm(pipeline, subpass, pCreateInfo->pColorBlendState,
+ pCreateInfo->pMultisampleState);
+ emit_3dstate_ps(pipeline, pCreateInfo->pColorBlendState);
#if GEN_GEN >= 8
- emit_3dstate_ps_extra(pipeline);
+ emit_3dstate_ps_extra(pipeline, subpass, pCreateInfo->pColorBlendState);
emit_3dstate_vf_topology(pipeline);
#endif
+ emit_3dstate_vf_statistics(pipeline);
*pPipeline = anv_pipeline_to_handle(pipeline);
- return VK_SUCCESS;
+ return pipeline->batch.status;
}
static VkResult
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
pipeline->device = device;
- pipeline->layout = anv_pipeline_layout_from_handle(pCreateInfo->layout);
pipeline->blend_state.map = NULL;
pipeline->batch.next = pipeline->batch.start = pipeline->batch_data;
pipeline->batch.end = pipeline->batch.start + sizeof(pipeline->batch_data);
pipeline->batch.relocs = &pipeline->batch_relocs;
+ pipeline->batch.status = VK_SUCCESS;
/* When we free the pipeline, we detect stages based on the NULL status
* of various prog_data pointers. Make them NULL by default.
pipeline->needs_data_cache = false;
assert(pCreateInfo->stage.stage == VK_SHADER_STAGE_COMPUTE_BIT);
+ pipeline->active_stages |= VK_SHADER_STAGE_COMPUTE_BIT;
ANV_FROM_HANDLE(anv_shader_module, module, pCreateInfo->stage.module);
result = anv_pipeline_compile_cs(pipeline, cache, pCreateInfo, module,
pCreateInfo->stage.pName,
const uint32_t subslices = MAX2(physical_device->subslice_total, 1);
+ const struct anv_shader_bin *cs_bin =
+ pipeline->shaders[MESA_SHADER_COMPUTE];
+
anv_batch_emit(&pipeline->batch, GENX(MEDIA_VFE_STATE), vfe) {
- vfe.ScratchSpaceBasePointer = (struct anv_address) {
- .bo = anv_scratch_pool_alloc(device, &device->scratch_pool,
- MESA_SHADER_COMPUTE,
- cs_prog_data->base.total_scratch),
- .offset = 0,
- };
- vfe.PerThreadScratchSpace = ffs(cs_prog_data->base.total_scratch / 2048);
#if GEN_GEN > 7
vfe.StackSize = 0;
#else
vfe.MaximumNumberofThreads =
devinfo->max_cs_threads * subslices - 1;
vfe.NumberofURBEntries = GEN_GEN <= 7 ? 0 : 2;
+#if GEN_GEN < 11
vfe.ResetGatewayTimer = true;
+#endif
#if GEN_GEN <= 8
vfe.BypassGatewayControl = true;
#endif
vfe.URBEntryAllocationSize = GEN_GEN <= 7 ? 0 : 2;
vfe.CURBEAllocationSize = vfe_curbe_allocation;
+
+ vfe.PerThreadScratchSpace = get_scratch_space(cs_bin);
+ vfe.ScratchSpaceBasePointer =
+ get_scratch_address(pipeline, MESA_SHADER_COMPUTE, cs_bin);
}
+ struct GENX(INTERFACE_DESCRIPTOR_DATA) desc = {
+ .KernelStartPointer = cs_bin->kernel.offset,
+
+ .SamplerCount = get_sampler_count(cs_bin),
+ .BindingTableEntryCount = get_binding_table_entry_count(cs_bin),
+ .BarrierEnable = cs_prog_data->uses_barrier,
+ .SharedLocalMemorySize =
+ encode_slm_size(GEN_GEN, cs_prog_data->base.total_shared),
+
+#if !GEN_IS_HASWELL
+ .ConstantURBEntryReadOffset = 0,
+#endif
+ .ConstantURBEntryReadLength = cs_prog_data->push.per_thread.regs,
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+ .CrossThreadConstantDataReadLength =
+ cs_prog_data->push.cross_thread.regs,
+#endif
+
+ .NumberofThreadsinGPGPUThreadGroup = cs_prog_data->threads,
+ };
+ GENX(INTERFACE_DESCRIPTOR_DATA_pack)(NULL,
+ pipeline->interface_descriptor_data,
+ &desc);
+
*pPipeline = anv_pipeline_to_handle(pipeline);
- return VK_SUCCESS;
+ return pipeline->batch.status;
}
VkResult genX(CreateGraphicsPipelines)(
VkResult result = VK_SUCCESS;
- unsigned i = 0;
- for (; i < count; i++) {
+ unsigned i;
+ for (i = 0; i < count; i++) {
result = genX(graphics_pipeline_create)(_device,
pipeline_cache,
&pCreateInfos[i],
pAllocator, &pPipelines[i]);
- if (result != VK_SUCCESS) {
- for (unsigned j = 0; j < i; j++) {
- anv_DestroyPipeline(_device, pPipelines[j], pAllocator);
- }
- return result;
- }
+ /* Bail out on the first error as it is not obvious what error should be
+ * report upon 2 different failures. */
+ if (result != VK_SUCCESS)
+ break;
}
- return VK_SUCCESS;
+ for (; i < count; i++)
+ pPipelines[i] = VK_NULL_HANDLE;
+
+ return result;
}
VkResult genX(CreateComputePipelines)(
VkResult result = VK_SUCCESS;
- unsigned i = 0;
- for (; i < count; i++) {
+ unsigned i;
+ for (i = 0; i < count; i++) {
result = compute_pipeline_create(_device, pipeline_cache,
&pCreateInfos[i],
pAllocator, &pPipelines[i]);
- if (result != VK_SUCCESS) {
- for (unsigned j = 0; j < i; j++) {
- anv_DestroyPipeline(_device, pPipelines[j], pAllocator);
- }
- return result;
- }
+ /* Bail out on the first error as it is not obvious what error should be
+ * report upon 2 different failures. */
+ if (result != VK_SUCCESS)
+ break;
}
- return VK_SUCCESS;
+ for (; i < count; i++)
+ pPipelines[i] = VK_NULL_HANDLE;
+
+ return result;
}