#include "anv_private.h"
#include "vk_format_info.h"
#include "vk_util.h"
+#include "util/fast_idiv_by_const.h"
#include "common/gen_l3_config.h"
#include "genxml/gen_macros.h"
#include "genxml/genX_pack.h"
-static void
-emit_lrm(struct anv_batch *batch, uint32_t reg, struct anv_address addr)
-{
- anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = reg;
- lrm.MemoryAddress = addr;
- }
-}
+/* We reserve GPR 14 and 15 for conditional rendering */
+#define GEN_MI_BUILDER_NUM_ALLOC_GPRS 14
+#define __gen_get_batch_dwords anv_batch_emit_dwords
+#define __gen_address_offset anv_address_add
+#include "common/gen_mi_builder.h"
static void
emit_lri(struct anv_batch *batch, uint32_t reg, uint32_t imm)
}
}
-#if GEN_IS_HASWELL || GEN_GEN >= 8
-static void
-emit_lrr(struct anv_batch *batch, uint32_t dst, uint32_t src)
-{
- anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_REG), lrr) {
- lrr.SourceRegisterAddress = src;
- lrr.DestinationRegisterAddress = dst;
- }
-}
-#endif
-
void
genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
{
sba.GeneralStateMOCS = GENX(MOCS);
sba.GeneralStateBaseAddressModifyEnable = true;
+ sba.StatelessDataPortAccessMOCS = GENX(MOCS);
+
sba.SurfaceStateBaseAddress =
anv_cmd_buffer_surface_base_address(cmd_buffer);
sba.SurfaceStateMOCS = GENX(MOCS);
sba.IndirectObjectBufferSizeModifyEnable = true;
sba.InstructionBufferSize = 0xfffff;
sba.InstructionBuffersizeModifyEnable = true;
+# else
+ /* On gen7, we have upper bounds instead. According to the docs,
+ * setting an upper bound of zero means that no bounds checking is
+ * performed so, in theory, we should be able to leave them zero.
+ * However, border color is broken and the GPU bounds-checks anyway.
+ * To avoid this and other potential problems, we may as well set it
+ * for everything.
+ */
+ sba.GeneralStateAccessUpperBound =
+ (struct anv_address) { .bo = NULL, .offset = 0xfffff000 };
+ sba.GeneralStateAccessUpperBoundModifyEnable = true;
+ sba.DynamicStateAccessUpperBound =
+ (struct anv_address) { .bo = NULL, .offset = 0xfffff000 };
+ sba.DynamicStateAccessUpperBoundModifyEnable = true;
+ sba.InstructionAccessUpperBound =
+ (struct anv_address) { .bo = NULL, .offset = 0xfffff000 };
+ sba.InstructionAccessUpperBoundModifyEnable = true;
# endif
# if (GEN_GEN >= 9)
- sba.BindlessSurfaceStateBaseAddress = (struct anv_address) { NULL, 0 };
+ if (cmd_buffer->device->instance->physicalDevice.use_softpin) {
+ sba.BindlessSurfaceStateBaseAddress = (struct anv_address) {
+ .bo = device->surface_state_pool.block_pool.bo,
+ .offset = 0,
+ };
+ sba.BindlessSurfaceStateSize = (1 << 20) - 1;
+ } else {
+ sba.BindlessSurfaceStateBaseAddress = ANV_NULL_ADDRESS;
+ sba.BindlessSurfaceStateSize = 0;
+ }
sba.BindlessSurfaceStateMOCS = GENX(MOCS);
sba.BindlessSurfaceStateBaseAddressModifyEnable = true;
- sba.BindlessSurfaceStateSize = 0;
# endif
# if (GEN_GEN >= 10)
sba.BindlessSamplerStateBaseAddress = (struct anv_address) { NULL, 0 };
union isl_color_value *fast_clear_color)
{
struct anv_attachment_state *att_state = &cmd_state->attachments[att];
- struct anv_image_view *iview = cmd_state->framebuffer->attachments[att];
+ struct anv_image_view *iview = cmd_state->attachments[att].image_view;
assert(iview->n_planes == 1);
struct anv_render_pass_attachment *pass_att =
&cmd_state->pass->attachments[att];
struct anv_attachment_state *att_state = &cmd_state->attachments[att];
- struct anv_image_view *iview = cmd_state->framebuffer->attachments[att];
+ struct anv_image_view *iview = cmd_state->attachments[att].image_view;
/* These will be initialized after the first subpass transition. */
att_state->aux_usage = ISL_AUX_USAGE_NONE;
0, 0, 1, hiz_op);
}
-#define MI_PREDICATE_SRC0 0x2400
-#define MI_PREDICATE_SRC1 0x2408
+static inline bool
+vk_image_layout_stencil_write_optimal(VkImageLayout layout)
+{
+ return layout == VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL ||
+ layout == VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL;
+}
+
+/* Transitions a HiZ-enabled depth buffer from one layout to another. Unless
+ * the initial layout is undefined, the HiZ buffer and depth buffer will
+ * represent the same data at the end of this operation.
+ */
+static void
+transition_stencil_buffer(struct anv_cmd_buffer *cmd_buffer,
+ const struct anv_image *image,
+ uint32_t base_level, uint32_t level_count,
+ uint32_t base_layer, uint32_t layer_count,
+ VkImageLayout initial_layout,
+ VkImageLayout final_layout)
+{
+#if GEN_GEN == 7
+ uint32_t plane = anv_image_aspect_to_plane(image->aspects,
+ VK_IMAGE_ASPECT_STENCIL_BIT);
+
+ /* On gen7, we have to store a texturable version of the stencil buffer in
+ * a shadow whenever VK_IMAGE_USAGE_SAMPLED_BIT is set and copy back and
+ * forth at strategic points. Stencil writes are only allowed in three
+ * layouts:
+ *
+ * - VK_IMAGE_LAYOUT_GENERAL
+ * - VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
+ * - VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
+ * - VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
+ *
+ * For general, we have no nice opportunity to transition so we do the copy
+ * to the shadow unconditionally at the end of the subpass. For transfer
+ * destinations, we can update it as part of the transfer op. For the
+ * other two, we delay the copy until a transition into some other layout.
+ */
+ if (image->planes[plane].shadow_surface.isl.size_B > 0 &&
+ vk_image_layout_stencil_write_optimal(initial_layout) &&
+ !vk_image_layout_stencil_write_optimal(final_layout)) {
+ anv_image_copy_to_shadow(cmd_buffer, image,
+ VK_IMAGE_ASPECT_STENCIL_BIT,
+ base_level, level_count,
+ base_layer, layer_count);
+ }
+#endif /* GEN_GEN == 7 */
+}
+
+#define MI_PREDICATE_SRC0 0x2400
+#define MI_PREDICATE_SRC1 0x2408
+#define MI_PREDICATE_RESULT 0x2418
static void
set_image_compressed_bit(struct anv_cmd_buffer *cmd_buffer,
set_image_compressed_bit(cmd_buffer, image, aspect, 0, 0, 1, true);
}
-#if GEN_IS_HASWELL || GEN_GEN >= 8
-static inline uint32_t
-mi_alu(uint32_t opcode, uint32_t operand1, uint32_t operand2)
-{
- struct GENX(MI_MATH_ALU_INSTRUCTION) instr = {
- .ALUOpcode = opcode,
- .Operand1 = operand1,
- .Operand2 = operand2,
- };
-
- uint32_t dw;
- GENX(MI_MATH_ALU_INSTRUCTION_pack)(NULL, &dw, &instr);
-
- return dw;
-}
-#endif
-
-#define CS_GPR(n) (0x2600 + (n) * 8)
-
/* This is only really practical on haswell and above because it requires
* MI math in order to get it correct.
*/
enum isl_aux_op resolve_op,
enum anv_fast_clear_type fast_clear_supported)
{
- struct anv_address fast_clear_type_addr =
- anv_image_get_fast_clear_type_addr(cmd_buffer->device, image, aspect);
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
- /* Name some registers */
- const int image_fc_reg = MI_ALU_REG0;
- const int fc_imm_reg = MI_ALU_REG1;
- const int pred_reg = MI_ALU_REG2;
-
- uint32_t *dw;
+ const struct gen_mi_value fast_clear_type =
+ gen_mi_mem32(anv_image_get_fast_clear_type_addr(cmd_buffer->device,
+ image, aspect));
if (resolve_op == ISL_AUX_OP_FULL_RESOLVE) {
/* In this case, we're doing a full resolve which means we want the
* if the first slice has been fast-cleared, it is also marked as
* compressed. See also set_image_fast_clear_state.
*/
- struct anv_address compression_state_addr =
- anv_image_get_compression_state_addr(cmd_buffer->device, image,
- aspect, level, array_layer);
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = MI_PREDICATE_SRC0;
- lrm.MemoryAddress = compression_state_addr;
- }
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
- sdi.Address = compression_state_addr;
- sdi.ImmediateData = 0;
- }
+ const struct gen_mi_value compression_state =
+ gen_mi_mem32(anv_image_get_compression_state_addr(cmd_buffer->device,
+ image, aspect,
+ level, array_layer));
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC0),
+ compression_state);
+ gen_mi_store(&b, compression_state, gen_mi_imm(0));
if (level == 0 && array_layer == 0) {
/* If the predicate is true, we want to write 0 to the fast clear type
*
* clear_type = clear_type & ~predicate;
*/
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = CS_GPR(image_fc_reg);
- lrm.MemoryAddress = fast_clear_type_addr;
- }
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_REG), lrr) {
- lrr.DestinationRegisterAddress = CS_GPR(pred_reg);
- lrr.SourceRegisterAddress = MI_PREDICATE_SRC0;
- }
-
- dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
- dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, image_fc_reg);
- dw[2] = mi_alu(MI_ALU_LOADINV, MI_ALU_SRCB, pred_reg);
- dw[3] = mi_alu(MI_ALU_AND, 0, 0);
- dw[4] = mi_alu(MI_ALU_STORE, image_fc_reg, MI_ALU_ACCU);
-
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
- srm.MemoryAddress = fast_clear_type_addr;
- srm.RegisterAddress = CS_GPR(image_fc_reg);
- }
+ struct gen_mi_value new_fast_clear_type =
+ gen_mi_iand(&b, fast_clear_type,
+ gen_mi_inot(&b, gen_mi_reg64(MI_PREDICATE_SRC0)));
+ gen_mi_store(&b, fast_clear_type, new_fast_clear_type);
}
} else if (level == 0 && array_layer == 0) {
/* In this case, we are doing a partial resolve to get rid of fast-clear
assert(resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE);
assert(fast_clear_supported < ANV_FAST_CLEAR_ANY);
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = CS_GPR(image_fc_reg);
- lrm.MemoryAddress = fast_clear_type_addr;
- }
- emit_lri(&cmd_buffer->batch, CS_GPR(image_fc_reg) + 4, 0);
-
- emit_lri(&cmd_buffer->batch, CS_GPR(fc_imm_reg), fast_clear_supported);
- emit_lri(&cmd_buffer->batch, CS_GPR(fc_imm_reg) + 4, 0);
-
- /* We need to compute (fast_clear_supported < image->fast_clear).
- * We do this by subtracting and storing the carry bit.
- */
- dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
- dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, fc_imm_reg);
- dw[2] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCB, image_fc_reg);
- dw[3] = mi_alu(MI_ALU_SUB, 0, 0);
- dw[4] = mi_alu(MI_ALU_STORE, pred_reg, MI_ALU_CF);
-
- /* Store the predicate */
- emit_lrr(&cmd_buffer->batch, MI_PREDICATE_SRC0, CS_GPR(pred_reg));
+ /* We need to compute (fast_clear_supported < image->fast_clear) */
+ struct gen_mi_value pred =
+ gen_mi_ult(&b, gen_mi_imm(fast_clear_supported), fast_clear_type);
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC0),
+ gen_mi_value_ref(&b, pred));
/* If the predicate is true, we want to write 0 to the fast clear type
* and, if it's false, leave it alone. We can do this by writing
*
* clear_type = clear_type & ~predicate;
*/
- dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
- dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, image_fc_reg);
- dw[2] = mi_alu(MI_ALU_LOADINV, MI_ALU_SRCB, pred_reg);
- dw[3] = mi_alu(MI_ALU_AND, 0, 0);
- dw[4] = mi_alu(MI_ALU_STORE, image_fc_reg, MI_ALU_ACCU);
-
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
- srm.RegisterAddress = CS_GPR(image_fc_reg);
- srm.MemoryAddress = fast_clear_type_addr;
- }
+ struct gen_mi_value new_fast_clear_type =
+ gen_mi_iand(&b, fast_clear_type, gen_mi_inot(&b, pred));
+ gen_mi_store(&b, fast_clear_type, new_fast_clear_type);
} else {
/* In this case, we're trying to do a partial resolve on a slice that
* doesn't have clear color. There's nothing to do.
return;
}
- /* We use the first half of src0 for the actual predicate. Set the second
- * half of src0 and all of src1 to 0 as the predicate operation will be
- * doing an implicit src0 != src1.
- */
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC0 + 4, 0);
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1 , 0);
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1 + 4, 0);
+ /* Set src1 to 0 and use a != condition */
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC1), gen_mi_imm(0));
anv_batch_emit(&cmd_buffer->batch, GENX(MI_PREDICATE), mip) {
mip.LoadOperation = LOAD_LOADINV;
enum isl_aux_op resolve_op,
enum anv_fast_clear_type fast_clear_supported)
{
- struct anv_address fast_clear_type_addr =
- anv_image_get_fast_clear_type_addr(cmd_buffer->device, image, aspect);
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
+ struct gen_mi_value fast_clear_type_mem =
+ gen_mi_mem32(anv_image_get_fast_clear_type_addr(cmd_buffer->device,
+ image, aspect));
/* This only works for partial resolves and only when the clear color is
* all or nothing. On the upside, this emits less command streamer code
* can't sample from CCS surfaces. It's enough to just load the fast clear
* state into the predicate register.
*/
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
- lrm.RegisterAddress = MI_PREDICATE_SRC0;
- lrm.MemoryAddress = fast_clear_type_addr;
- }
- anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
- sdi.Address = fast_clear_type_addr;
- sdi.ImmediateData = 0;
- }
-
- /* We use the first half of src0 for the actual predicate. Set the second
- * half of src0 and all of src1 to 0 as the predicate operation will be
- * doing an implicit src0 != src1.
- */
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC0 + 4, 0);
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1 , 0);
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1 + 4, 0);
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC0), fast_clear_type_mem);
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC1), gen_mi_imm(0));
+ gen_mi_store(&b, fast_clear_type_mem, gen_mi_imm(0));
anv_batch_emit(&cmd_buffer->batch, GENX(MI_PREDICATE), mip) {
mip.LoadOperation = LOAD_LOADINV;
set_image_fast_clear_state(cmd_buffer, image, aspect,
ANV_FAST_CLEAR_NONE);
- /* The fast clear value dword(s) will be copied into a surface state object.
- * Ensure that the restrictions of the fields in the dword(s) are followed.
- *
- * CCS buffers on SKL+ can have any value set for the clear colors.
- */
- if (image->samples == 1 && GEN_GEN >= 9)
- return;
-
- /* Other combinations of auxiliary buffers and platforms require specific
- * values in the clear value dword(s).
+ /* Initialize the struct fields that are accessed for fast-clears so that
+ * the HW restrictions on the field values are satisfied.
*/
struct anv_address addr =
anv_image_get_clear_color_addr(cmd_buffer->device, image, aspect);
if (GEN_GEN >= 9) {
- for (unsigned i = 0; i < 4; i++) {
+ const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
+ const unsigned num_dwords = GEN_GEN >= 10 ?
+ isl_dev->ss.clear_color_state_size / 4 :
+ isl_dev->ss.clear_value_size / 4;
+ for (unsigned i = 0; i < num_dwords; i++) {
anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
sdi.Address = addr;
sdi.Address.offset += i * 4;
- /* MCS buffers on SKL+ can only have 1/0 clear colors. */
- assert(image->samples > 1);
sdi.ImmediateData = 0;
}
}
anv_image_get_clear_color_addr(cmd_buffer->device, image, aspect);
unsigned copy_size = cmd_buffer->device->isl_dev.ss.clear_value_size;
+#if GEN_GEN == 7
+ /* On gen7, the combination of commands used here(MI_LOAD_REGISTER_MEM
+ * and MI_STORE_REGISTER_MEM) can cause GPU hangs if any rendering is
+ * in-flight when they are issued even if the memory touched is not
+ * currently active for rendering. The weird bit is that it is not the
+ * MI_LOAD/STORE_REGISTER_MEM commands which hang but rather the in-flight
+ * rendering hangs such that the next stalling command after the
+ * MI_LOAD/STORE_REGISTER_MEM commands will catch the hang.
+ *
+ * It is unclear exactly why this hang occurs. Both MI commands come with
+ * warnings about the 3D pipeline but that doesn't seem to fully explain
+ * it. My (Jason's) best theory is that it has something to do with the
+ * fact that we're using a GPU state register as our temporary and that
+ * something with reading/writing it is causing problems.
+ *
+ * In order to work around this issue, we emit a PIPE_CONTROL with the
+ * command streamer stall bit set.
+ */
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+#endif
+
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
if (copy_from_surface_state) {
- genX(cmd_buffer_mi_memcpy)(cmd_buffer, entry_addr,
- ss_clear_addr, copy_size);
+ gen_mi_memcpy(&b, entry_addr, ss_clear_addr, copy_size);
} else {
- genX(cmd_buffer_mi_memcpy)(cmd_buffer, ss_clear_addr,
- entry_addr, copy_size);
+ gen_mi_memcpy(&b, ss_clear_addr, entry_addr, copy_size);
/* Updating a surface state object may require that the state cache be
* invalidated. From the SKL PRM, Shared Functions -> State -> State
assert(level_count != VK_REMAINING_MIP_LEVELS &&
layer_count != VK_REMAINING_ARRAY_LAYERS);
/* Ensure the subresource range is valid. */
- uint64_t last_level_num = base_level + level_count;
+ UNUSED uint64_t last_level_num = base_level + level_count;
const uint32_t max_depth = anv_minify(image->extent.depth, base_level);
UNUSED const uint32_t image_layers = MAX2(image->array_size, max_depth);
assert((uint64_t)base_layer + layer_count <= image_layers);
assert(isl_format_is_compressed(image->planes[plane].surface.isl.format));
assert(plane == 0);
anv_image_copy_to_shadow(cmd_buffer, image,
+ VK_IMAGE_ASPECT_COLOR_BIT,
base_level, level_count,
base_layer, layer_count);
}
{
const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
struct anv_cmd_state *state = &cmd_buffer->state;
+ struct anv_framebuffer *framebuffer = cmd_buffer->state.framebuffer;
vk_free(&cmd_buffer->pool->alloc, state->attachments);
next_state.offset += ss_stride;
next_state.map += ss_stride;
+ const VkRenderPassAttachmentBeginInfoKHR *begin_attachment =
+ vk_find_struct_const(begin, RENDER_PASS_ATTACHMENT_BEGIN_INFO_KHR);
+
+ if (begin && !begin_attachment)
+ assert(pass->attachment_count == framebuffer->attachment_count);
+
for (uint32_t i = 0; i < pass->attachment_count; ++i) {
if (vk_format_is_color(pass->attachments[i].format)) {
state->attachments[i].color.state = next_state;
next_state.offset += ss_stride;
next_state.map += ss_stride;
}
+
+ if (begin_attachment && begin_attachment->attachmentCount != 0) {
+ assert(begin_attachment->attachmentCount == pass->attachment_count);
+ ANV_FROM_HANDLE(anv_image_view, iview, begin_attachment->pAttachments[i]);
+ cmd_buffer->state.attachments[i].image_view = iview;
+ } else if (framebuffer && i < framebuffer->attachment_count) {
+ cmd_buffer->state.attachments[i].image_view = framebuffer->attachments[i];
+ }
}
assert(next_state.offset == state->render_pass_states.offset +
state->render_pass_states.alloc_size);
if (begin) {
- ANV_FROM_HANDLE(anv_framebuffer, framebuffer, begin->framebuffer);
- assert(pass->attachment_count == framebuffer->attachment_count);
-
isl_null_fill_state(isl_dev, state->null_surface_state.map,
isl_extent3d(framebuffer->width,
framebuffer->height,
if (clear_aspects)
state->attachments[i].clear_value = begin->pClearValues[i];
- struct anv_image_view *iview = framebuffer->attachments[i];
+ struct anv_image_view *iview = cmd_buffer->state.attachments[i].image_view;
anv_assert(iview->vk_format == att->format);
const uint32_t num_layers = iview->planes[0].isl.array_len;
cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;
}
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+ if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY) {
+ const VkCommandBufferInheritanceConditionalRenderingInfoEXT *conditional_rendering_info =
+ vk_find_struct_const(pBeginInfo->pInheritanceInfo->pNext, COMMAND_BUFFER_INHERITANCE_CONDITIONAL_RENDERING_INFO_EXT);
+
+ /* If secondary buffer supports conditional rendering
+ * we should emit commands as if conditional rendering is enabled.
+ */
+ cmd_buffer->state.conditional_render_enabled =
+ conditional_rendering_info && conditional_rendering_info->conditionalRenderingEnable;
+ }
+#endif
+
return result;
}
assert(secondary->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY);
assert(!anv_batch_has_error(&secondary->batch));
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+ if (secondary->state.conditional_render_enabled) {
+ if (!primary->state.conditional_render_enabled) {
+ /* Secondary buffer is constructed as if it will be executed
+ * with conditional rendering, we should satisfy this dependency
+ * regardless of conditional rendering being enabled in primary.
+ */
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &primary->batch);
+ gen_mi_store(&b, gen_mi_reg64(ANV_PREDICATE_RESULT_REG),
+ gen_mi_imm(UINT64_MAX));
+ }
+ }
+#endif
+
if (secondary->usage_flags &
VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
/* If we're continuing a render pass from the primary, we need to
assert(!urb_low_bw || cfg->n[GEN_L3P_URB] == cfg->n[GEN_L3P_SLM]);
/* Minimum number of ways that can be allocated to the URB. */
- MAYBE_UNUSED const unsigned n0_urb = devinfo->is_baytrail ? 32 : 0;
+ const unsigned n0_urb = devinfo->is_baytrail ? 32 : 0;
assert(cfg->n[GEN_L3P_URB] >= n0_urb);
uint32_t l3sqcr1, l3cr2, l3cr3;
* saying that render target writes are ongoing.
*/
if (bits & ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT)
- bits &= ~(ANV_PIPE_RENDER_TARGET_WRITES);
+ bits &= ~(ANV_PIPE_RENDER_TARGET_BUFFER_WRITES);
bits &= ~(ANV_PIPE_FLUSH_BITS | ANV_PIPE_CS_STALL_BIT);
}
const VkImageSubresourceRange *range =
&pImageMemoryBarriers[i].subresourceRange;
+ uint32_t base_layer, layer_count;
+ if (image->type == VK_IMAGE_TYPE_3D) {
+ base_layer = 0;
+ layer_count = anv_minify(image->extent.depth, range->baseMipLevel);
+ } else {
+ base_layer = range->baseArrayLayer;
+ layer_count = anv_get_layerCount(image, range);
+ }
+
if (range->aspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) {
transition_depth_buffer(cmd_buffer, image,
pImageMemoryBarriers[i].oldLayout,
pImageMemoryBarriers[i].newLayout);
- } else if (range->aspectMask & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
+ }
+
+ if (range->aspectMask & VK_IMAGE_ASPECT_STENCIL_BIT) {
+ transition_stencil_buffer(cmd_buffer, image,
+ range->baseMipLevel,
+ anv_get_levelCount(image, range),
+ base_layer, layer_count,
+ pImageMemoryBarriers[i].oldLayout,
+ pImageMemoryBarriers[i].newLayout);
+ }
+
+ if (range->aspectMask & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
VkImageAspectFlags color_aspects =
anv_image_expand_aspects(image, range->aspectMask);
uint32_t aspect_bit;
-
- uint32_t base_layer, layer_count;
- if (image->type == VK_IMAGE_TYPE_3D) {
- base_layer = 0;
- layer_count = anv_minify(image->extent.depth, range->baseMipLevel);
- } else {
- base_layer = range->baseArrayLayer;
- layer_count = anv_get_layerCount(image, range);
- }
-
anv_foreach_image_aspect_bit(aspect_bit, image, color_aspects) {
transition_color_buffer(cmd_buffer, image, 1UL << aspect_bit,
range->baseMipLevel,
return pipe_state->dynamic_offsets[dynamic_offset_idx];
}
+static struct anv_address
+anv_descriptor_set_address(struct anv_cmd_buffer *cmd_buffer,
+ struct anv_descriptor_set *set)
+{
+ if (set->pool) {
+ /* This is a normal descriptor set */
+ return (struct anv_address) {
+ .bo = &set->pool->bo,
+ .offset = set->desc_mem.offset,
+ };
+ } else {
+ /* This is a push descriptor set. We have to flag it as used on the GPU
+ * so that the next time we push descriptors, we grab a new memory.
+ */
+ struct anv_push_descriptor_set *push_set =
+ (struct anv_push_descriptor_set *)set;
+ push_set->set_used_on_gpu = true;
+
+ return (struct anv_address) {
+ .bo = cmd_buffer->dynamic_state_stream.state_pool->block_pool.bo,
+ .offset = set->desc_mem.offset,
+ };
+ }
+}
+
static VkResult
emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
gl_shader_stage stage,
struct anv_state *bt_state)
{
- const struct gen_device_info *devinfo = &cmd_buffer->device->info;
struct anv_subpass *subpass = cmd_buffer->state.subpass;
struct anv_cmd_pipeline_state *pipe_state;
struct anv_pipeline *pipeline;
- uint32_t bias, state_offset;
+ uint32_t state_offset;
switch (stage) {
case MESA_SHADER_COMPUTE:
pipe_state = &cmd_buffer->state.compute.base;
- bias = 1;
break;
default:
pipe_state = &cmd_buffer->state.gfx.base;
- bias = 0;
break;
}
pipeline = pipe_state->pipeline;
}
struct anv_pipeline_bind_map *map = &pipeline->shaders[stage]->bind_map;
- if (bias + map->surface_count == 0) {
+ if (map->surface_count == 0) {
*bt_state = (struct anv_state) { 0, };
return VK_SUCCESS;
}
*bt_state = anv_cmd_buffer_alloc_binding_table(cmd_buffer,
- bias + map->surface_count,
+ map->surface_count,
&state_offset);
uint32_t *bt_map = bt_state->map;
if (bt_state->map == NULL)
return VK_ERROR_OUT_OF_DEVICE_MEMORY;
- if (stage == MESA_SHADER_COMPUTE &&
- get_cs_prog_data(pipeline)->uses_num_work_groups) {
- struct anv_state surface_state;
- surface_state =
- anv_cmd_buffer_alloc_surface_state(cmd_buffer);
-
- const enum isl_format format =
- anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER);
- anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
- format,
- cmd_buffer->state.compute.num_workgroups,
- 12, 1);
-
- bt_map[0] = surface_state.offset + state_offset;
- add_surface_reloc(cmd_buffer, surface_state,
- cmd_buffer->state.compute.num_workgroups);
- }
-
- if (map->surface_count == 0)
- goto out;
-
- /* We only use push constant space for images before gen9 */
- if (map->image_count > 0 && devinfo->gen < 9) {
- VkResult result =
- anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, images);
- if (result != VK_SUCCESS)
- return result;
-
- cmd_buffer->state.push_constants_dirty |= 1 << stage;
- }
+ /* We only need to emit relocs if we're not using softpin. If we are using
+ * softpin then we always keep all user-allocated memory objects resident.
+ */
+ const bool need_client_mem_relocs =
+ !cmd_buffer->device->instance->physicalDevice.use_softpin;
- uint32_t image = 0;
for (uint32_t s = 0; s < map->surface_count; s++) {
struct anv_pipeline_binding *binding = &map->surface_to_descriptor[s];
surface_state = cmd_buffer->state.null_surface_state;
}
- bt_map[bias + s] = surface_state.offset + state_offset;
+ bt_map[s] = surface_state.offset + state_offset;
continue;
} else if (binding->set == ANV_DESCRIPTOR_SET_SHADER_CONSTANTS) {
struct anv_state surface_state =
surface_state, format,
constant_data, constant_data_size, 1);
- bt_map[bias + s] = surface_state.offset + state_offset;
+ bt_map[s] = surface_state.offset + state_offset;
add_surface_reloc(cmd_buffer, surface_state, constant_data);
continue;
+ } else if (binding->set == ANV_DESCRIPTOR_SET_NUM_WORK_GROUPS) {
+ /* This is always the first binding for compute shaders */
+ assert(stage == MESA_SHADER_COMPUTE && s == 0);
+ if (!get_cs_prog_data(pipeline)->uses_num_work_groups)
+ continue;
+
+ struct anv_state surface_state =
+ anv_cmd_buffer_alloc_surface_state(cmd_buffer);
+
+ const enum isl_format format =
+ anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER);
+ anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
+ format,
+ cmd_buffer->state.compute.num_workgroups,
+ 12, 1);
+ bt_map[s] = surface_state.offset + state_offset;
+ if (need_client_mem_relocs) {
+ add_surface_reloc(cmd_buffer, surface_state,
+ cmd_buffer->state.compute.num_workgroups);
+ }
+ continue;
+ } else if (binding->set == ANV_DESCRIPTOR_SET_DESCRIPTORS) {
+ /* This is a descriptor set buffer so the set index is actually
+ * given by binding->binding. (Yes, that's confusing.)
+ */
+ struct anv_descriptor_set *set =
+ pipe_state->descriptors[binding->binding];
+ assert(set->desc_mem.alloc_size);
+ assert(set->desc_surface_state.alloc_size);
+ bt_map[s] = set->desc_surface_state.offset + state_offset;
+ add_surface_reloc(cmd_buffer, set->desc_surface_state,
+ anv_descriptor_set_address(cmd_buffer, set));
+ continue;
}
const struct anv_descriptor *desc =
desc->image_view->planes[binding->plane].optimal_sampler_surface_state;
surface_state = sstate.state;
assert(surface_state.alloc_size);
- add_surface_state_relocs(cmd_buffer, sstate);
+ if (need_client_mem_relocs)
+ add_surface_state_relocs(cmd_buffer, sstate);
break;
}
case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
desc->image_view->planes[binding->plane].optimal_sampler_surface_state;
surface_state = sstate.state;
assert(surface_state.alloc_size);
- add_surface_state_relocs(cmd_buffer, sstate);
+ if (need_client_mem_relocs)
+ add_surface_state_relocs(cmd_buffer, sstate);
} else {
/* For color input attachments, we create the surface state at
* vkBeginRenderPass time so that we can include aux and clear
: desc->image_view->planes[binding->plane].storage_surface_state;
surface_state = sstate.state;
assert(surface_state.alloc_size);
- add_surface_state_relocs(cmd_buffer, sstate);
- if (devinfo->gen < 9) {
- assert(image < MAX_GEN8_IMAGES);
- struct brw_image_param *image_param =
- &cmd_buffer->state.push_constants[stage]->images[image];
-
- *image_param =
- desc->image_view->planes[binding->plane].storage_image_param;
- }
- image++;
+ if (need_client_mem_relocs)
+ add_surface_state_relocs(cmd_buffer, sstate);
break;
}
case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
surface_state = desc->buffer_view->surface_state;
assert(surface_state.alloc_size);
- add_surface_reloc(cmd_buffer, surface_state,
- desc->buffer_view->address);
+ if (need_client_mem_relocs) {
+ add_surface_reloc(cmd_buffer, surface_state,
+ desc->buffer_view->address);
+ }
break;
case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
format, address, range, 1);
- add_surface_reloc(cmd_buffer, surface_state, address);
+ if (need_client_mem_relocs)
+ add_surface_reloc(cmd_buffer, surface_state, address);
break;
}
? desc->buffer_view->writeonly_storage_surface_state
: desc->buffer_view->storage_surface_state;
assert(surface_state.alloc_size);
- add_surface_reloc(cmd_buffer, surface_state,
- desc->buffer_view->address);
- if (devinfo->gen < 9) {
- assert(image < MAX_GEN8_IMAGES);
- struct brw_image_param *image_param =
- &cmd_buffer->state.push_constants[stage]->images[image];
-
- *image_param = desc->buffer_view->storage_image_param;
+ if (need_client_mem_relocs) {
+ add_surface_reloc(cmd_buffer, surface_state,
+ desc->buffer_view->address);
}
- image++;
break;
default:
continue;
}
- bt_map[bias + s] = surface_state.offset + state_offset;
+ bt_map[s] = surface_state.offset + state_offset;
}
- assert(image == map->image_count);
- out:
#if GEN_GEN >= 11
/* The PIPE_CONTROL command description says:
*
DIV_ROUND_UP(constant_data_size, 32) - range->start);
read_addr = anv_address_add(constant_data,
range->start * 32);
+ } else if (binding->set == ANV_DESCRIPTOR_SET_DESCRIPTORS) {
+ /* This is a descriptor set buffer so the set index is
+ * actually given by binding->binding. (Yes, that's
+ * confusing.)
+ */
+ struct anv_descriptor_set *set =
+ gfx_state->base.descriptors[binding->binding];
+ struct anv_address desc_buffer_addr =
+ anv_descriptor_set_address(cmd_buffer, set);
+ const unsigned desc_buffer_size = set->desc_mem.alloc_size;
+
+ read_len = MIN2(range->length,
+ DIV_ROUND_UP(desc_buffer_size, 32) - range->start);
+ read_addr = anv_address_add(desc_buffer_addr,
+ range->start * 32);
} else {
const struct anv_descriptor *desc =
anv_descriptor_for_binding(&gfx_state->base, binding);
cmd_buffer->state.gfx.vb_dirty &= ~vb_emit;
+#if GEN_GEN >= 8
+ if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_XFB_ENABLE) {
+ /* We don't need any per-buffer dirty tracking because you're not
+ * allowed to bind different XFB buffers while XFB is enabled.
+ */
+ for (unsigned idx = 0; idx < MAX_XFB_BUFFERS; idx++) {
+ struct anv_xfb_binding *xfb = &cmd_buffer->state.xfb_bindings[idx];
+ anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_SO_BUFFER), sob) {
+ sob.SOBufferIndex = idx;
+
+ if (cmd_buffer->state.xfb_enabled && xfb->buffer && xfb->size != 0) {
+ sob.SOBufferEnable = true;
+ sob.MOCS = cmd_buffer->device->default_mocs,
+ sob.StreamOffsetWriteEnable = false;
+ sob.SurfaceBaseAddress = anv_address_add(xfb->buffer->address,
+ xfb->offset);
+ /* Size is in DWords - 1 */
+ sob.SurfaceSize = xfb->size / 4 - 1;
+ }
+ }
+ }
+
+ /* CNL and later require a CS stall after 3DSTATE_SO_BUFFER */
+ if (GEN_GEN >= 10)
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
+ }
+#endif
+
if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_PIPELINE) {
anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);
genX(cmd_buffer_flush_state)(cmd_buffer);
+ if (cmd_buffer->state.conditional_render_enabled)
+ genX(cmd_emit_conditional_render_predicate)(cmd_buffer);
+
if (vs_prog_data->uses_firstvertex ||
vs_prog_data->uses_baseinstance)
emit_base_vertex_instance(cmd_buffer, firstVertex, firstInstance);
instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
+ prim.PredicateEnable = cmd_buffer->state.conditional_render_enabled;
prim.VertexAccessType = SEQUENTIAL;
prim.PrimitiveTopologyType = pipeline->topology;
prim.VertexCountPerInstance = vertexCount;
prim.StartInstanceLocation = firstInstance;
prim.BaseVertexLocation = 0;
}
-
- cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_RENDER_TARGET_WRITES;
}
void genX(CmdDrawIndexed)(
genX(cmd_buffer_flush_state)(cmd_buffer);
+ if (cmd_buffer->state.conditional_render_enabled)
+ genX(cmd_emit_conditional_render_predicate)(cmd_buffer);
+
if (vs_prog_data->uses_firstvertex ||
vs_prog_data->uses_baseinstance)
emit_base_vertex_instance(cmd_buffer, vertexOffset, firstInstance);
instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
+ prim.PredicateEnable = cmd_buffer->state.conditional_render_enabled;
prim.VertexAccessType = RANDOM;
prim.PrimitiveTopologyType = pipeline->topology;
prim.VertexCountPerInstance = indexCount;
prim.StartInstanceLocation = firstInstance;
prim.BaseVertexLocation = vertexOffset;
}
-
- cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_RENDER_TARGET_WRITES;
}
/* Auto-Draw / Indirect Registers */
#define GEN7_3DPRIM_START_INSTANCE 0x243C
#define GEN7_3DPRIM_BASE_VERTEX 0x2440
-/* MI_MATH only exists on Haswell+ */
-#if GEN_IS_HASWELL || GEN_GEN >= 8
-
-/* Emit dwords to multiply GPR0 by N */
-static void
-build_alu_multiply_gpr0(uint32_t *dw, unsigned *dw_count, uint32_t N)
+void genX(CmdDrawIndirectByteCountEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t instanceCount,
+ uint32_t firstInstance,
+ VkBuffer counterBuffer,
+ VkDeviceSize counterBufferOffset,
+ uint32_t counterOffset,
+ uint32_t vertexStride)
{
- VK_OUTARRAY_MAKE(out, dw, dw_count);
+#if GEN_IS_HASWELL || GEN_GEN >= 8
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+ ANV_FROM_HANDLE(anv_buffer, counter_buffer, counterBuffer);
+ struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
+ const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);
-#define append_alu(opcode, operand1, operand2) \
- vk_outarray_append(&out, alu_dw) *alu_dw = mi_alu(opcode, operand1, operand2)
+ /* firstVertex is always zero for this draw function */
+ const uint32_t firstVertex = 0;
- assert(N > 0);
- unsigned top_bit = 31 - __builtin_clz(N);
- for (int i = top_bit - 1; i >= 0; i--) {
- /* We get our initial data in GPR0 and we write the final data out to
- * GPR0 but we use GPR1 as our scratch register.
- */
- unsigned src_reg = i == top_bit - 1 ? MI_ALU_REG0 : MI_ALU_REG1;
- unsigned dst_reg = i == 0 ? MI_ALU_REG0 : MI_ALU_REG1;
-
- /* Shift the current value left by 1 */
- append_alu(MI_ALU_LOAD, MI_ALU_SRCA, src_reg);
- append_alu(MI_ALU_LOAD, MI_ALU_SRCB, src_reg);
- append_alu(MI_ALU_ADD, 0, 0);
-
- if (N & (1 << i)) {
- /* Store ACCU to R1 and add R0 to R1 */
- append_alu(MI_ALU_STORE, MI_ALU_REG1, MI_ALU_ACCU);
- append_alu(MI_ALU_LOAD, MI_ALU_SRCA, MI_ALU_REG0);
- append_alu(MI_ALU_LOAD, MI_ALU_SRCB, MI_ALU_REG1);
- append_alu(MI_ALU_ADD, 0, 0);
- }
+ if (anv_batch_has_error(&cmd_buffer->batch))
+ return;
- append_alu(MI_ALU_STORE, dst_reg, MI_ALU_ACCU);
- }
+ genX(cmd_buffer_flush_state)(cmd_buffer);
-#undef append_alu
-}
+ if (vs_prog_data->uses_firstvertex ||
+ vs_prog_data->uses_baseinstance)
+ emit_base_vertex_instance(cmd_buffer, firstVertex, firstInstance);
+ if (vs_prog_data->uses_drawid)
+ emit_draw_index(cmd_buffer, 0);
-static void
-emit_mul_gpr0(struct anv_batch *batch, uint32_t N)
-{
- uint32_t num_dwords;
- build_alu_multiply_gpr0(NULL, &num_dwords, N);
+ /* Our implementation of VK_KHR_multiview uses instancing to draw the
+ * different views. We need to multiply instanceCount by the view count.
+ */
+ instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);
- uint32_t *dw = anv_batch_emitn(batch, 1 + num_dwords, GENX(MI_MATH));
- build_alu_multiply_gpr0(dw + 1, &num_dwords, N);
-}
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+ struct gen_mi_value count =
+ gen_mi_mem32(anv_address_add(counter_buffer->address,
+ counterBufferOffset));
+ if (counterOffset)
+ count = gen_mi_isub(&b, count, gen_mi_imm(counterOffset));
+ count = gen_mi_udiv32_imm(&b, count, vertexStride);
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_VERTEX_COUNT), count);
+
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_START_VERTEX),
+ gen_mi_imm(firstVertex));
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_INSTANCE_COUNT),
+ gen_mi_imm(instanceCount));
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE),
+ gen_mi_imm(firstInstance));
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX), gen_mi_imm(0));
+ anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
+ prim.IndirectParameterEnable = true;
+ prim.VertexAccessType = SEQUENTIAL;
+ prim.PrimitiveTopologyType = pipeline->topology;
+ }
#endif /* GEN_IS_HASWELL || GEN_GEN >= 8 */
+}
static void
load_indirect_parameters(struct anv_cmd_buffer *cmd_buffer,
struct anv_address addr,
bool indexed)
{
- struct anv_batch *batch = &cmd_buffer->batch;
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
- emit_lrm(batch, GEN7_3DPRIM_VERTEX_COUNT, anv_address_add(addr, 0));
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_VERTEX_COUNT),
+ gen_mi_mem32(anv_address_add(addr, 0)));
+ struct gen_mi_value instance_count = gen_mi_mem32(anv_address_add(addr, 4));
unsigned view_count = anv_subpass_view_count(cmd_buffer->state.subpass);
if (view_count > 1) {
#if GEN_IS_HASWELL || GEN_GEN >= 8
- emit_lrm(batch, CS_GPR(0), anv_address_add(addr, 4));
- emit_mul_gpr0(batch, view_count);
- emit_lrr(batch, GEN7_3DPRIM_INSTANCE_COUNT, CS_GPR(0));
+ instance_count = gen_mi_imul_imm(&b, instance_count, view_count);
#else
anv_finishme("Multiview + indirect draw requires MI_MATH; "
"MI_MATH is not supported on Ivy Bridge");
- emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, anv_address_add(addr, 4));
#endif
- } else {
- emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, anv_address_add(addr, 4));
}
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_INSTANCE_COUNT), instance_count);
- emit_lrm(batch, GEN7_3DPRIM_START_VERTEX, anv_address_add(addr, 8));
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_START_VERTEX),
+ gen_mi_mem32(anv_address_add(addr, 8)));
if (indexed) {
- emit_lrm(batch, GEN7_3DPRIM_BASE_VERTEX, anv_address_add(addr, 12));
- emit_lrm(batch, GEN7_3DPRIM_START_INSTANCE, anv_address_add(addr, 16));
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX),
+ gen_mi_mem32(anv_address_add(addr, 12)));
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE),
+ gen_mi_mem32(anv_address_add(addr, 16)));
} else {
- emit_lrm(batch, GEN7_3DPRIM_START_INSTANCE, anv_address_add(addr, 12));
- emit_lri(batch, GEN7_3DPRIM_BASE_VERTEX, 0);
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_START_INSTANCE),
+ gen_mi_mem32(anv_address_add(addr, 12)));
+ gen_mi_store(&b, gen_mi_reg32(GEN7_3DPRIM_BASE_VERTEX), gen_mi_imm(0));
}
}
genX(cmd_buffer_flush_state)(cmd_buffer);
+ if (cmd_buffer->state.conditional_render_enabled)
+ genX(cmd_emit_conditional_render_predicate)(cmd_buffer);
+
for (uint32_t i = 0; i < drawCount; i++) {
struct anv_address draw = anv_address_add(buffer->address, offset);
anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
prim.IndirectParameterEnable = true;
+ prim.PredicateEnable = cmd_buffer->state.conditional_render_enabled;
prim.VertexAccessType = SEQUENTIAL;
prim.PrimitiveTopologyType = pipeline->topology;
}
offset += stride;
}
-
- cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_RENDER_TARGET_WRITES;
}
void genX(CmdDrawIndexedIndirect)(
genX(cmd_buffer_flush_state)(cmd_buffer);
+ if (cmd_buffer->state.conditional_render_enabled)
+ genX(cmd_emit_conditional_render_predicate)(cmd_buffer);
+
for (uint32_t i = 0; i < drawCount; i++) {
struct anv_address draw = anv_address_add(buffer->address, offset);
anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
prim.IndirectParameterEnable = true;
+ prim.PredicateEnable = cmd_buffer->state.conditional_render_enabled;
prim.VertexAccessType = RANDOM;
prim.PrimitiveTopologyType = pipeline->topology;
}
offset += stride;
}
-
- cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_RENDER_TARGET_WRITES;
}
+#define TMP_DRAW_COUNT_REG 0x2670 /* MI_ALU_REG14 */
+
static void
prepare_for_draw_count_predicate(struct anv_cmd_buffer *cmd_buffer,
- struct anv_address count_address)
+ struct anv_address count_address,
+ const bool conditional_render_enabled)
{
- /* Upload the current draw count from the draw parameters buffer to
- * MI_PREDICATE_SRC0.
- */
- emit_lrm(&cmd_buffer->batch, MI_PREDICATE_SRC0, count_address);
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC0 + 4, 0);
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1 + 4, 0);
+ if (conditional_render_enabled) {
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+ gen_mi_store(&b, gen_mi_reg64(TMP_DRAW_COUNT_REG),
+ gen_mi_mem32(count_address));
+#endif
+ } else {
+ /* Upload the current draw count from the draw parameters buffer to
+ * MI_PREDICATE_SRC0.
+ */
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC0),
+ gen_mi_mem32(count_address));
+
+ gen_mi_store(&b, gen_mi_reg32(MI_PREDICATE_SRC1 + 4), gen_mi_imm(0));
+ }
}
static void
emit_draw_count_predicate(struct anv_cmd_buffer *cmd_buffer,
uint32_t draw_index)
{
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
/* Upload the index of the current primitive to MI_PREDICATE_SRC1. */
- emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1, draw_index);
+ gen_mi_store(&b, gen_mi_reg32(MI_PREDICATE_SRC1), gen_mi_imm(draw_index));
if (draw_index == 0) {
anv_batch_emit(&cmd_buffer->batch, GENX(MI_PREDICATE), mip) {
}
}
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+static void
+emit_draw_count_predicate_with_conditional_render(
+ struct anv_cmd_buffer *cmd_buffer,
+ uint32_t draw_index)
+{
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
+ struct gen_mi_value pred = gen_mi_ult(&b, gen_mi_imm(draw_index),
+ gen_mi_reg64(TMP_DRAW_COUNT_REG));
+ pred = gen_mi_iand(&b, pred, gen_mi_reg64(ANV_PREDICATE_RESULT_REG));
+
+#if GEN_GEN >= 8
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_RESULT), pred);
+#else
+ /* MI_PREDICATE_RESULT is not whitelisted in i915 command parser
+ * so we emit MI_PREDICATE to set it.
+ */
+
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC0), pred);
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC1), gen_mi_imm(0));
+
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_PREDICATE), mip) {
+ mip.LoadOperation = LOAD_LOADINV;
+ mip.CombineOperation = COMBINE_SET;
+ mip.CompareOperation = COMPARE_SRCS_EQUAL;
+ }
+#endif
+}
+#endif
+
void genX(CmdDrawIndirectCountKHR)(
VkCommandBuffer commandBuffer,
VkBuffer _buffer,
struct anv_address count_address =
anv_address_add(count_buffer->address, countBufferOffset);
- prepare_for_draw_count_predicate(cmd_buffer, count_address);
+ prepare_for_draw_count_predicate(cmd_buffer, count_address,
+ cmd_state->conditional_render_enabled);
for (uint32_t i = 0; i < maxDrawCount; i++) {
struct anv_address draw = anv_address_add(buffer->address, offset);
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+ if (cmd_state->conditional_render_enabled) {
+ emit_draw_count_predicate_with_conditional_render(cmd_buffer, i);
+ } else {
+ emit_draw_count_predicate(cmd_buffer, i);
+ }
+#else
emit_draw_count_predicate(cmd_buffer, i);
+#endif
if (vs_prog_data->uses_firstvertex ||
vs_prog_data->uses_baseinstance)
offset += stride;
}
-
- cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_RENDER_TARGET_WRITES;
}
void genX(CmdDrawIndexedIndirectCountKHR)(
struct anv_address count_address =
anv_address_add(count_buffer->address, countBufferOffset);
- prepare_for_draw_count_predicate(cmd_buffer, count_address);
+ prepare_for_draw_count_predicate(cmd_buffer, count_address,
+ cmd_state->conditional_render_enabled);
for (uint32_t i = 0; i < maxDrawCount; i++) {
struct anv_address draw = anv_address_add(buffer->address, offset);
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+ if (cmd_state->conditional_render_enabled) {
+ emit_draw_count_predicate_with_conditional_render(cmd_buffer, i);
+ } else {
+ emit_draw_count_predicate(cmd_buffer, i);
+ }
+#else
emit_draw_count_predicate(cmd_buffer, i);
+#endif
/* TODO: We need to stomp base vertex to 0 somehow */
if (vs_prog_data->uses_firstvertex ||
offset += stride;
}
+}
+
+void genX(CmdBeginTransformFeedbackEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t firstCounterBuffer,
+ uint32_t counterBufferCount,
+ const VkBuffer* pCounterBuffers,
+ const VkDeviceSize* pCounterBufferOffsets)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+
+ assert(firstCounterBuffer < MAX_XFB_BUFFERS);
+ assert(counterBufferCount <= MAX_XFB_BUFFERS);
+ assert(firstCounterBuffer + counterBufferCount <= MAX_XFB_BUFFERS);
+
+ /* From the SKL PRM Vol. 2c, SO_WRITE_OFFSET:
+ *
+ * "Ssoftware must ensure that no HW stream output operations can be in
+ * process or otherwise pending at the point that the MI_LOAD/STORE
+ * commands are processed. This will likely require a pipeline flush."
+ */
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
- cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_RENDER_TARGET_WRITES;
+ for (uint32_t idx = 0; idx < MAX_XFB_BUFFERS; idx++) {
+ /* If we have a counter buffer, this is a resume so we need to load the
+ * value into the streamout offset register. Otherwise, this is a begin
+ * and we need to reset it to zero.
+ */
+ if (pCounterBuffers &&
+ idx >= firstCounterBuffer &&
+ idx - firstCounterBuffer < counterBufferCount &&
+ pCounterBuffers[idx - firstCounterBuffer] != VK_NULL_HANDLE) {
+ uint32_t cb_idx = idx - firstCounterBuffer;
+ ANV_FROM_HANDLE(anv_buffer, counter_buffer, pCounterBuffers[cb_idx]);
+ uint64_t offset = pCounterBufferOffsets ?
+ pCounterBufferOffsets[cb_idx] : 0;
+
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
+ lrm.RegisterAddress = GENX(SO_WRITE_OFFSET0_num) + idx * 4;
+ lrm.MemoryAddress = anv_address_add(counter_buffer->address,
+ offset);
+ }
+ } else {
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
+ lri.RegisterOffset = GENX(SO_WRITE_OFFSET0_num) + idx * 4;
+ lri.DataDWord = 0;
+ }
+ }
+ }
+
+ cmd_buffer->state.xfb_enabled = true;
+ cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_XFB_ENABLE;
+}
+
+void genX(CmdEndTransformFeedbackEXT)(
+ VkCommandBuffer commandBuffer,
+ uint32_t firstCounterBuffer,
+ uint32_t counterBufferCount,
+ const VkBuffer* pCounterBuffers,
+ const VkDeviceSize* pCounterBufferOffsets)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+
+ assert(firstCounterBuffer < MAX_XFB_BUFFERS);
+ assert(counterBufferCount <= MAX_XFB_BUFFERS);
+ assert(firstCounterBuffer + counterBufferCount <= MAX_XFB_BUFFERS);
+
+ /* From the SKL PRM Vol. 2c, SO_WRITE_OFFSET:
+ *
+ * "Ssoftware must ensure that no HW stream output operations can be in
+ * process or otherwise pending at the point that the MI_LOAD/STORE
+ * commands are processed. This will likely require a pipeline flush."
+ */
+ cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
+ for (uint32_t cb_idx = 0; cb_idx < counterBufferCount; cb_idx++) {
+ unsigned idx = firstCounterBuffer + cb_idx;
+
+ /* If we have a counter buffer, this is a resume so we need to load the
+ * value into the streamout offset register. Otherwise, this is a begin
+ * and we need to reset it to zero.
+ */
+ if (pCounterBuffers &&
+ cb_idx < counterBufferCount &&
+ pCounterBuffers[cb_idx] != VK_NULL_HANDLE) {
+ ANV_FROM_HANDLE(anv_buffer, counter_buffer, pCounterBuffers[cb_idx]);
+ uint64_t offset = pCounterBufferOffsets ?
+ pCounterBufferOffsets[cb_idx] : 0;
+
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
+ srm.MemoryAddress = anv_address_add(counter_buffer->address,
+ offset);
+ srm.RegisterAddress = GENX(SO_WRITE_OFFSET0_num) + idx * 4;
+ }
+ }
+ }
+
+ cmd_buffer->state.xfb_enabled = false;
+ cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_XFB_ENABLE;
}
static VkResult
genX(cmd_buffer_flush_compute_state)(struct anv_cmd_buffer *cmd_buffer)
{
struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
- MAYBE_UNUSED VkResult result;
+ VkResult result;
assert(pipeline->active_stages == VK_SHADER_STAGE_COMPUTE_BIT);
if (anv_batch_has_error(&cmd_buffer->batch))
return;
- VkResult result =
- anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, MESA_SHADER_COMPUTE,
- base_work_group_id);
- if (result != VK_SUCCESS) {
- cmd_buffer->batch.status = result;
- return;
- }
-
struct anv_push_constants *push =
- cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
+ &cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
if (push->base_work_group_id[0] != baseGroupX ||
push->base_work_group_id[1] != baseGroupY ||
push->base_work_group_id[2] != baseGroupZ) {
genX(cmd_buffer_flush_compute_state)(cmd_buffer);
+ if (cmd_buffer->state.conditional_render_enabled)
+ genX(cmd_emit_conditional_render_predicate)(cmd_buffer);
+
anv_batch_emit(&cmd_buffer->batch, GENX(GPGPU_WALKER), ggw) {
+ ggw.PredicateEnable = cmd_buffer->state.conditional_render_enabled;
ggw.SIMDSize = prog_data->simd_size / 16;
ggw.ThreadDepthCounterMaximum = 0;
ggw.ThreadHeightCounterMaximum = 0;
genX(cmd_buffer_flush_compute_state)(cmd_buffer);
- emit_lrm(batch, GPGPU_DISPATCHDIMX, anv_address_add(addr, 0));
- emit_lrm(batch, GPGPU_DISPATCHDIMY, anv_address_add(addr, 4));
- emit_lrm(batch, GPGPU_DISPATCHDIMZ, anv_address_add(addr, 8));
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
-#if GEN_GEN <= 7
- /* Clear upper 32-bits of SRC0 and all 64-bits of SRC1 */
- emit_lri(batch, MI_PREDICATE_SRC0 + 4, 0);
- emit_lri(batch, MI_PREDICATE_SRC1 + 0, 0);
- emit_lri(batch, MI_PREDICATE_SRC1 + 4, 0);
+ struct gen_mi_value size_x = gen_mi_mem32(anv_address_add(addr, 0));
+ struct gen_mi_value size_y = gen_mi_mem32(anv_address_add(addr, 4));
+ struct gen_mi_value size_z = gen_mi_mem32(anv_address_add(addr, 8));
- /* Load compute_dispatch_indirect_x_size into SRC0 */
- emit_lrm(batch, MI_PREDICATE_SRC0, anv_address_add(addr, 0));
+ gen_mi_store(&b, gen_mi_reg32(GPGPU_DISPATCHDIMX), size_x);
+ gen_mi_store(&b, gen_mi_reg32(GPGPU_DISPATCHDIMY), size_y);
+ gen_mi_store(&b, gen_mi_reg32(GPGPU_DISPATCHDIMZ), size_z);
+#if GEN_GEN <= 7
/* predicate = (compute_dispatch_indirect_x_size == 0); */
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC0), size_x);
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC1), gen_mi_imm(0));
anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
mip.LoadOperation = LOAD_LOAD;
mip.CombineOperation = COMBINE_SET;
mip.CompareOperation = COMPARE_SRCS_EQUAL;
}
- /* Load compute_dispatch_indirect_y_size into SRC0 */
- emit_lrm(batch, MI_PREDICATE_SRC0, anv_address_add(addr, 4));
-
/* predicate |= (compute_dispatch_indirect_y_size == 0); */
+ gen_mi_store(&b, gen_mi_reg32(MI_PREDICATE_SRC0), size_y);
anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
mip.LoadOperation = LOAD_LOAD;
mip.CombineOperation = COMBINE_OR;
mip.CompareOperation = COMPARE_SRCS_EQUAL;
}
- /* Load compute_dispatch_indirect_z_size into SRC0 */
- emit_lrm(batch, MI_PREDICATE_SRC0, anv_address_add(addr, 8));
-
/* predicate |= (compute_dispatch_indirect_z_size == 0); */
+ gen_mi_store(&b, gen_mi_reg32(MI_PREDICATE_SRC0), size_z);
anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
mip.LoadOperation = LOAD_LOAD;
mip.CombineOperation = COMBINE_OR;
}
/* predicate = !predicate; */
-#define COMPARE_FALSE 1
anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
mip.LoadOperation = LOAD_LOADINV;
mip.CombineOperation = COMBINE_OR;
mip.CompareOperation = COMPARE_FALSE;
}
+
+#if GEN_IS_HASWELL
+ if (cmd_buffer->state.conditional_render_enabled) {
+ /* predicate &= !(conditional_rendering_predicate == 0); */
+ gen_mi_store(&b, gen_mi_reg32(MI_PREDICATE_SRC0),
+ gen_mi_reg32(ANV_PREDICATE_RESULT_REG));
+ anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
+ mip.LoadOperation = LOAD_LOADINV;
+ mip.CombineOperation = COMBINE_AND;
+ mip.CompareOperation = COMPARE_SRCS_EQUAL;
+ }
+ }
+#endif
+
+#else /* GEN_GEN > 7 */
+ if (cmd_buffer->state.conditional_render_enabled)
+ genX(cmd_emit_conditional_render_predicate)(cmd_buffer);
#endif
anv_batch_emit(batch, GENX(GPGPU_WALKER), ggw) {
ggw.IndirectParameterEnable = true;
- ggw.PredicateEnable = GEN_GEN <= 7;
+ ggw.PredicateEnable = GEN_GEN <= 7 ||
+ cmd_buffer->state.conditional_render_enabled;
ggw.SIMDSize = prog_data->simd_size / 16;
ggw.ThreadDepthCounterMaximum = 0;
ggw.ThreadHeightCounterMaximum = 0;
anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_CC_STATE_POINTERS), t);
#endif
+#if GEN_GEN == 9
+ if (pipeline == _3D) {
+ /* There is a mid-object preemption workaround which requires you to
+ * re-emit MEDIA_VFE_STATE after switching from GPGPU to 3D. However,
+ * even without preemption, we have issues with geometry flickering when
+ * GPGPU and 3D are back-to-back and this seems to fix it. We don't
+ * really know why.
+ */
+ const uint32_t subslices =
+ MAX2(cmd_buffer->device->instance->physicalDevice.subslice_total, 1);
+ anv_batch_emit(&cmd_buffer->batch, GENX(MEDIA_VFE_STATE), vfe) {
+ vfe.MaximumNumberofThreads =
+ devinfo->max_cs_threads * subslices - 1;
+ vfe.NumberofURBEntries = 2;
+ vfe.URBEntryAllocationSize = 2;
+ }
+ }
+#endif
+
/* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
* PIPELINE_SELECT [DevBWR+]":
*
assert(a < cmd_state->pass->attachment_count);
struct anv_attachment_state *att_state = &cmd_state->attachments[a];
- struct anv_image_view *iview = fb->attachments[a];
+ struct anv_image_view *iview = cmd_state->attachments[a].image_view;
const struct anv_image *image = iview->image;
/* A resolve is necessary before use as an input attachment if the clear
target_layout = subpass->attachments[i].layout;
}
+ uint32_t base_layer, layer_count;
+ if (image->type == VK_IMAGE_TYPE_3D) {
+ base_layer = 0;
+ layer_count = anv_minify(iview->image->extent.depth,
+ iview->planes[0].isl.base_level);
+ } else {
+ base_layer = iview->planes[0].isl.base_array_layer;
+ layer_count = fb->layers;
+ }
+
if (image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
assert(image->aspects == VK_IMAGE_ASPECT_COLOR_BIT);
-
- uint32_t base_layer, layer_count;
- if (image->type == VK_IMAGE_TYPE_3D) {
- base_layer = 0;
- layer_count = anv_minify(iview->image->extent.depth,
- iview->planes[0].isl.base_level);
- } else {
- base_layer = iview->planes[0].isl.base_array_layer;
- layer_count = fb->layers;
- }
-
transition_color_buffer(cmd_buffer, image, VK_IMAGE_ASPECT_COLOR_BIT,
iview->planes[0].isl.base_level, 1,
base_layer, layer_count,
att_state->current_layout, target_layout);
- } else if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
+ }
+
+ if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
transition_depth_buffer(cmd_buffer, image,
att_state->current_layout, target_layout);
att_state->aux_usage =
anv_layout_to_aux_usage(&cmd_buffer->device->info, image,
VK_IMAGE_ASPECT_DEPTH_BIT, target_layout);
}
+
+ if (image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
+ transition_stencil_buffer(cmd_buffer, image,
+ iview->planes[0].isl.base_level, 1,
+ base_layer, layer_count,
+ att_state->current_layout, target_layout);
+ }
att_state->current_layout = target_layout;
if (att_state->pending_clear_aspects & VK_IMAGE_ASPECT_COLOR_BIT) {
cmd_buffer->state.attachments[dst_att].pending_clear_aspects = 0;
}
- struct anv_image_view *src_iview = fb->attachments[src_att];
- struct anv_image_view *dst_iview = fb->attachments[dst_att];
+ struct anv_image_view *src_iview = cmd_state->attachments[src_att].image_view;
+ struct anv_image_view *dst_iview = cmd_state->attachments[dst_att].image_view;
const VkRect2D render_area = cmd_buffer->state.render_area;
cmd_buffer->state.attachments[dst_att].pending_clear_aspects = 0;
}
- struct anv_image_view *src_iview = fb->attachments[src_att];
- struct anv_image_view *dst_iview = fb->attachments[dst_att];
+ struct anv_image_view *src_iview = cmd_state->attachments[src_att].image_view;
+ struct anv_image_view *dst_iview = cmd_state->attachments[dst_att].image_view;
const VkRect2D render_area = cmd_buffer->state.render_area;
}
}
+#if GEN_GEN == 7
+ /* On gen7, we have to store a texturable version of the stencil buffer in
+ * a shadow whenever VK_IMAGE_USAGE_SAMPLED_BIT is set and copy back and
+ * forth at strategic points. Stencil writes are only allowed in three
+ * layouts:
+ *
+ * - VK_IMAGE_LAYOUT_GENERAL
+ * - VK_IMAGE_LAYOUT_TRANSFER_DST_OPTIMAL
+ * - VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL
+ * - VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL
+ *
+ * For general, we have no nice opportunity to transition so we do the copy
+ * to the shadow unconditionally at the end of the subpass. For transfer
+ * destinations, we can update it as part of the transfer op. For the
+ * other two, we delay the copy until a transition into some other layout.
+ */
+ if (subpass->depth_stencil_attachment) {
+ uint32_t a = subpass->depth_stencil_attachment->attachment;
+ assert(a != VK_ATTACHMENT_UNUSED);
+
+ struct anv_attachment_state *att_state = &cmd_state->attachments[a];
+ struct anv_image_view *iview = cmd_state->attachments[a].image_view;;
+ const struct anv_image *image = iview->image;
+
+ if (image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
+ uint32_t plane = anv_image_aspect_to_plane(image->aspects,
+ VK_IMAGE_ASPECT_STENCIL_BIT);
+
+ if (image->planes[plane].shadow_surface.isl.size_B > 0 &&
+ att_state->current_layout == VK_IMAGE_LAYOUT_GENERAL) {
+ assert(image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT);
+ anv_image_copy_to_shadow(cmd_buffer, image,
+ VK_IMAGE_ASPECT_STENCIL_BIT,
+ iview->planes[plane].isl.base_level, 1,
+ iview->planes[plane].isl.base_array_layer,
+ fb->layers);
+ }
+ }
+ }
+#endif /* GEN_GEN == 7 */
+
for (uint32_t i = 0; i < subpass->attachment_count; ++i) {
const uint32_t a = subpass->attachments[i].attachment;
if (a == VK_ATTACHMENT_UNUSED)
assert(a < cmd_state->pass->attachment_count);
struct anv_attachment_state *att_state = &cmd_state->attachments[a];
- struct anv_image_view *iview = fb->attachments[a];
+ struct anv_image_view *iview = cmd_state->attachments[a].image_view;
const struct anv_image *image = iview->image;
if ((image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) &&
* SRGB view & a UNORM image).
*/
if (fast_clear_type != ANV_FAST_CLEAR_NONE) {
- anv_perf_warn(cmd_buffer->device->instance, fb,
+ anv_perf_warn(cmd_buffer->device->instance, iview,
"Doing a partial resolve to get rid of clear color at the "
"end of a renderpass due to an image/view format mismatch");
VkImageLayout target_layout =
cmd_state->pass->attachments[a].final_layout;
+ uint32_t base_layer, layer_count;
+ if (image->type == VK_IMAGE_TYPE_3D) {
+ base_layer = 0;
+ layer_count = anv_minify(iview->image->extent.depth,
+ iview->planes[0].isl.base_level);
+ } else {
+ base_layer = iview->planes[0].isl.base_array_layer;
+ layer_count = fb->layers;
+ }
+
if (image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
assert(image->aspects == VK_IMAGE_ASPECT_COLOR_BIT);
-
- uint32_t base_layer, layer_count;
- if (image->type == VK_IMAGE_TYPE_3D) {
- base_layer = 0;
- layer_count = anv_minify(iview->image->extent.depth,
- iview->planes[0].isl.base_level);
- } else {
- base_layer = iview->planes[0].isl.base_array_layer;
- layer_count = fb->layers;
- }
-
transition_color_buffer(cmd_buffer, image, VK_IMAGE_ASPECT_COLOR_BIT,
iview->planes[0].isl.base_level, 1,
base_layer, layer_count,
att_state->current_layout, target_layout);
- } else if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
+ }
+
+ if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
transition_depth_buffer(cmd_buffer, image,
att_state->current_layout, target_layout);
}
+
+ if (image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
+ transition_stencil_buffer(cmd_buffer, image,
+ iview->planes[0].isl.base_level, 1,
+ base_layer, layer_count,
+ att_state->current_layout, target_layout);
+ }
}
/* Accumulate any subpass flushes that need to happen after the subpass.
cmd_buffer->state.hiz_enabled = false;
#ifndef NDEBUG
- anv_dump_add_framebuffer(cmd_buffer, cmd_buffer->state.framebuffer);
+ anv_dump_add_attachments(cmd_buffer);
#endif
/* Remove references to render pass specific state. This enables us to
{
genX(CmdEndRenderPass)(commandBuffer);
}
+
+void
+genX(cmd_emit_conditional_render_predicate)(struct anv_cmd_buffer *cmd_buffer)
+{
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC0),
+ gen_mi_reg32(ANV_PREDICATE_RESULT_REG));
+ gen_mi_store(&b, gen_mi_reg64(MI_PREDICATE_SRC1), gen_mi_imm(0));
+
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_PREDICATE), mip) {
+ mip.LoadOperation = LOAD_LOADINV;
+ mip.CombineOperation = COMBINE_SET;
+ mip.CompareOperation = COMPARE_SRCS_EQUAL;
+ }
+#endif
+}
+
+#if GEN_GEN >= 8 || GEN_IS_HASWELL
+void genX(CmdBeginConditionalRenderingEXT)(
+ VkCommandBuffer commandBuffer,
+ const VkConditionalRenderingBeginInfoEXT* pConditionalRenderingBegin)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+ ANV_FROM_HANDLE(anv_buffer, buffer, pConditionalRenderingBegin->buffer);
+ struct anv_cmd_state *cmd_state = &cmd_buffer->state;
+ struct anv_address value_address =
+ anv_address_add(buffer->address, pConditionalRenderingBegin->offset);
+
+ const bool isInverted = pConditionalRenderingBegin->flags &
+ VK_CONDITIONAL_RENDERING_INVERTED_BIT_EXT;
+
+ cmd_state->conditional_render_enabled = true;
+
+ genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+
+ struct gen_mi_builder b;
+ gen_mi_builder_init(&b, &cmd_buffer->batch);
+
+ /* Section 19.4 of the Vulkan 1.1.85 spec says:
+ *
+ * If the value of the predicate in buffer memory changes
+ * while conditional rendering is active, the rendering commands
+ * may be discarded in an implementation-dependent way.
+ * Some implementations may latch the value of the predicate
+ * upon beginning conditional rendering while others
+ * may read it before every rendering command.
+ *
+ * So it's perfectly fine to read a value from the buffer once.
+ */
+ struct gen_mi_value value = gen_mi_mem32(value_address);
+
+ /* Precompute predicate result, it is necessary to support secondary
+ * command buffers since it is unknown if conditional rendering is
+ * inverted when populating them.
+ */
+ gen_mi_store(&b, gen_mi_reg64(ANV_PREDICATE_RESULT_REG),
+ isInverted ? gen_mi_uge(&b, gen_mi_imm(0), value) :
+ gen_mi_ult(&b, gen_mi_imm(0), value));
+}
+
+void genX(CmdEndConditionalRenderingEXT)(
+ VkCommandBuffer commandBuffer)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+ struct anv_cmd_state *cmd_state = &cmd_buffer->state;
+
+ cmd_state->conditional_render_enabled = false;
+}
+#endif
+
+/* Set of stage bits for which are pipelined, i.e. they get queued by the
+ * command streamer for later execution.
+ */
+#define ANV_PIPELINE_STAGE_PIPELINED_BITS \
+ (VK_PIPELINE_STAGE_VERTEX_INPUT_BIT | \
+ VK_PIPELINE_STAGE_VERTEX_SHADER_BIT | \
+ VK_PIPELINE_STAGE_TESSELLATION_CONTROL_SHADER_BIT | \
+ VK_PIPELINE_STAGE_TESSELLATION_EVALUATION_SHADER_BIT | \
+ VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT | \
+ VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT | \
+ VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT | \
+ VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT | \
+ VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT | \
+ VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT | \
+ VK_PIPELINE_STAGE_TRANSFER_BIT | \
+ VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT | \
+ VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT | \
+ VK_PIPELINE_STAGE_ALL_COMMANDS_BIT)
+
+void genX(CmdSetEvent)(
+ VkCommandBuffer commandBuffer,
+ VkEvent _event,
+ VkPipelineStageFlags stageMask)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+ ANV_FROM_HANDLE(anv_event, event, _event);
+
+ anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
+ if (stageMask & ANV_PIPELINE_STAGE_PIPELINED_BITS) {
+ pc.StallAtPixelScoreboard = true;
+ pc.CommandStreamerStallEnable = true;
+ }
+
+ pc.DestinationAddressType = DAT_PPGTT,
+ pc.PostSyncOperation = WriteImmediateData,
+ pc.Address = (struct anv_address) {
+ cmd_buffer->device->dynamic_state_pool.block_pool.bo,
+ event->state.offset
+ };
+ pc.ImmediateData = VK_EVENT_SET;
+ }
+}
+
+void genX(CmdResetEvent)(
+ VkCommandBuffer commandBuffer,
+ VkEvent _event,
+ VkPipelineStageFlags stageMask)
+{
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+ ANV_FROM_HANDLE(anv_event, event, _event);
+
+ anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
+ if (stageMask & ANV_PIPELINE_STAGE_PIPELINED_BITS) {
+ pc.StallAtPixelScoreboard = true;
+ pc.CommandStreamerStallEnable = true;
+ }
+
+ pc.DestinationAddressType = DAT_PPGTT;
+ pc.PostSyncOperation = WriteImmediateData;
+ pc.Address = (struct anv_address) {
+ cmd_buffer->device->dynamic_state_pool.block_pool.bo,
+ event->state.offset
+ };
+ pc.ImmediateData = VK_EVENT_RESET;
+ }
+}
+
+void genX(CmdWaitEvents)(
+ VkCommandBuffer commandBuffer,
+ uint32_t eventCount,
+ const VkEvent* pEvents,
+ VkPipelineStageFlags srcStageMask,
+ VkPipelineStageFlags destStageMask,
+ uint32_t memoryBarrierCount,
+ const VkMemoryBarrier* pMemoryBarriers,
+ uint32_t bufferMemoryBarrierCount,
+ const VkBufferMemoryBarrier* pBufferMemoryBarriers,
+ uint32_t imageMemoryBarrierCount,
+ const VkImageMemoryBarrier* pImageMemoryBarriers)
+{
+#if GEN_GEN >= 8
+ ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
+
+ for (uint32_t i = 0; i < eventCount; i++) {
+ ANV_FROM_HANDLE(anv_event, event, pEvents[i]);
+
+ anv_batch_emit(&cmd_buffer->batch, GENX(MI_SEMAPHORE_WAIT), sem) {
+ sem.WaitMode = PollingMode,
+ sem.CompareOperation = COMPARE_SAD_EQUAL_SDD,
+ sem.SemaphoreDataDword = VK_EVENT_SET,
+ sem.SemaphoreAddress = (struct anv_address) {
+ cmd_buffer->device->dynamic_state_pool.block_pool.bo,
+ event->state.offset
+ };
+ }
+ }
+#else
+ anv_finishme("Implement events on gen7");
+#endif
+
+ genX(CmdPipelineBarrier)(commandBuffer, srcStageMask, destStageMask,
+ false, /* byRegion */
+ memoryBarrierCount, pMemoryBarriers,
+ bufferMemoryBarrierCount, pBufferMemoryBarriers,
+ imageMemoryBarrierCount, pImageMemoryBarriers);
+}