#include "tu_private.h"
-#include "registers/adreno_pm4.xml.h"
-#include "registers/adreno_common.xml.h"
+#include "adreno_pm4.xml.h"
+#include "adreno_common.xml.h"
#include "vk_format.h"
#include "tu_cs.h"
-#include "tu_blit.h"
-
-#define OVERFLOW_FLAG_REG REG_A6XX_CP_SCRATCH_REG(0)
void
tu_bo_list_init(struct tu_bo_list *list)
return VK_SUCCESS;
}
-static bool
-is_linear_mipmapped(const struct tu_image_view *iview)
-{
- return iview->image->layout.tile_mode == TILE6_LINEAR &&
- iview->base_mip != iview->image->level_count - 1;
-}
-
-static bool
-force_sysmem(const struct tu_cmd_buffer *cmd,
- const struct VkRect2D *render_area)
-{
- const struct tu_framebuffer *fb = cmd->state.framebuffer;
- const struct tu_physical_device *device = cmd->device->physical_device;
- bool has_linear_mipmapped_store = false;
- const struct tu_render_pass *pass = cmd->state.pass;
-
- /* Layered rendering requires sysmem. */
- if (fb->layers > 1)
- return true;
-
- /* Iterate over all the places we call tu6_emit_store_attachment() */
- for (unsigned i = 0; i < pass->subpass_count; i++) {
- const struct tu_subpass *subpass = &pass->subpasses[i];
- if (subpass->resolve_attachments) {
- for (unsigned i = 0; i < subpass->color_count; i++) {
- uint32_t a = subpass->resolve_attachments[i].attachment;
- if (a != VK_ATTACHMENT_UNUSED &&
- cmd->state.pass->attachments[a].store_op == VK_ATTACHMENT_STORE_OP_STORE) {
- const struct tu_image_view *iview = fb->attachments[a].attachment;
- if (is_linear_mipmapped(iview)) {
- has_linear_mipmapped_store = true;
- break;
- }
- }
- }
- }
+void
+tu6_emit_event_write(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs,
+ enum vgt_event_type event)
+{
+ bool need_seqno = false;
+ switch (event) {
+ case CACHE_FLUSH_TS:
+ case WT_DONE_TS:
+ case RB_DONE_TS:
+ case PC_CCU_FLUSH_DEPTH_TS:
+ case PC_CCU_FLUSH_COLOR_TS:
+ case PC_CCU_RESOLVE_TS:
+ need_seqno = true;
+ break;
+ default:
+ break;
}
- for (unsigned i = 0; i < pass->attachment_count; i++) {
- if (pass->attachments[i].gmem_offset >= 0 &&
- cmd->state.pass->attachments[i].store_op == VK_ATTACHMENT_STORE_OP_STORE) {
- const struct tu_image_view *iview = fb->attachments[i].attachment;
- if (is_linear_mipmapped(iview)) {
- has_linear_mipmapped_store = true;
- break;
- }
- }
+ tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, need_seqno ? 4 : 1);
+ tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(event));
+ if (need_seqno) {
+ tu_cs_emit_qw(cs, global_iova(cmd, seqno_dummy));
+ tu_cs_emit(cs, 0);
}
-
- /* Linear textures cannot have any padding between mipmap levels and their
- * height isn't padded, while at the same time the GMEM->MEM resolve does
- * not have per-pixel granularity, so if the image height isn't aligned to
- * the resolve granularity and the render area is tall enough, we may wind
- * up writing past the bottom of the image into the next miplevel or even
- * past the end of the image. For the last miplevel, the layout code should
- * insert enough padding so that the overdraw writes to the padding. To
- * work around this, we force-enable sysmem rendering.
- */
- const uint32_t y2 = render_area->offset.y + render_area->extent.height;
- const uint32_t aligned_y2 = ALIGN_POT(y2, device->tile_align_h);
-
- return has_linear_mipmapped_store && aligned_y2 > fb->height;
}
static void
-tu_tiling_config_update_tile_layout(struct tu_tiling_config *tiling,
- const struct tu_device *dev,
- uint32_t pixels)
-{
- const uint32_t tile_align_w = dev->physical_device->tile_align_w;
- const uint32_t tile_align_h = dev->physical_device->tile_align_h;
- const uint32_t max_tile_width = 1024; /* A6xx */
-
- /* note: don't offset the tiling config by render_area.offset,
- * because binning pass can't deal with it
- * this means we might end up with more tiles than necessary,
- * but load/store/etc are still scissored to the render_area
+tu6_emit_flushes(struct tu_cmd_buffer *cmd_buffer,
+ struct tu_cs *cs,
+ enum tu_cmd_flush_bits flushes)
+{
+ /* Experiments show that invalidating CCU while it still has data in it
+ * doesn't work, so make sure to always flush before invalidating in case
+ * any data remains that hasn't yet been made available through a barrier.
+ * However it does seem to work for UCHE.
*/
- tiling->tile0.offset = (VkOffset2D) {};
-
- const uint32_t ra_width =
- tiling->render_area.extent.width +
- (tiling->render_area.offset.x - tiling->tile0.offset.x);
- const uint32_t ra_height =
- tiling->render_area.extent.height +
- (tiling->render_area.offset.y - tiling->tile0.offset.y);
-
- /* start from 1 tile */
- tiling->tile_count = (VkExtent2D) {
- .width = 1,
- .height = 1,
- };
- tiling->tile0.extent = (VkExtent2D) {
- .width = align(ra_width, tile_align_w),
- .height = align(ra_height, tile_align_h),
- };
-
- if (unlikely(dev->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN)) {
- /* start with 2x2 tiles */
- tiling->tile_count.width = 2;
- tiling->tile_count.height = 2;
- tiling->tile0.extent.width = align(DIV_ROUND_UP(ra_width, 2), tile_align_w);
- tiling->tile0.extent.height = align(DIV_ROUND_UP(ra_height, 2), tile_align_h);
- }
-
- /* do not exceed max tile width */
- while (tiling->tile0.extent.width > max_tile_width) {
- tiling->tile_count.width++;
- tiling->tile0.extent.width =
- align(DIV_ROUND_UP(ra_width, tiling->tile_count.width), tile_align_w);
- }
-
- /* will force to sysmem, don't bother trying to have a valid tile config
- * TODO: just skip all GMEM stuff when sysmem is forced?
- */
- if (!pixels)
- return;
-
- /* do not exceed gmem size */
- while (tiling->tile0.extent.width * tiling->tile0.extent.height > pixels) {
- if (tiling->tile0.extent.width > MAX2(tile_align_w, tiling->tile0.extent.height)) {
- tiling->tile_count.width++;
- tiling->tile0.extent.width =
- align(DIV_ROUND_UP(ra_width, tiling->tile_count.width), tile_align_w);
- } else {
- /* if this assert fails then layout is impossible.. */
- assert(tiling->tile0.extent.height > tile_align_h);
- tiling->tile_count.height++;
- tiling->tile0.extent.height =
- align(DIV_ROUND_UP(ra_height, tiling->tile_count.height), tile_align_h);
- }
- }
+ if (flushes & (TU_CMD_FLAG_CCU_FLUSH_COLOR |
+ TU_CMD_FLAG_CCU_INVALIDATE_COLOR))
+ tu6_emit_event_write(cmd_buffer, cs, PC_CCU_FLUSH_COLOR_TS);
+ if (flushes & (TU_CMD_FLAG_CCU_FLUSH_DEPTH |
+ TU_CMD_FLAG_CCU_INVALIDATE_DEPTH))
+ tu6_emit_event_write(cmd_buffer, cs, PC_CCU_FLUSH_DEPTH_TS);
+ if (flushes & TU_CMD_FLAG_CCU_INVALIDATE_COLOR)
+ tu6_emit_event_write(cmd_buffer, cs, PC_CCU_INVALIDATE_COLOR);
+ if (flushes & TU_CMD_FLAG_CCU_INVALIDATE_DEPTH)
+ tu6_emit_event_write(cmd_buffer, cs, PC_CCU_INVALIDATE_DEPTH);
+ if (flushes & TU_CMD_FLAG_CACHE_FLUSH)
+ tu6_emit_event_write(cmd_buffer, cs, CACHE_FLUSH_TS);
+ if (flushes & TU_CMD_FLAG_CACHE_INVALIDATE)
+ tu6_emit_event_write(cmd_buffer, cs, CACHE_INVALIDATE);
+ if (flushes & TU_CMD_FLAG_WAIT_MEM_WRITES)
+ tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
+ if (flushes & TU_CMD_FLAG_WAIT_FOR_IDLE)
+ tu_cs_emit_wfi(cs);
+ if (flushes & TU_CMD_FLAG_WAIT_FOR_ME)
+ tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
}
-static void
-tu_tiling_config_update_pipe_layout(struct tu_tiling_config *tiling,
- const struct tu_device *dev)
-{
- const uint32_t max_pipe_count = 32; /* A6xx */
-
- /* start from 1 tile per pipe */
- tiling->pipe0 = (VkExtent2D) {
- .width = 1,
- .height = 1,
- };
- tiling->pipe_count = tiling->tile_count;
-
- while (tiling->pipe_count.width * tiling->pipe_count.height > max_pipe_count) {
- if (tiling->pipe0.width < tiling->pipe0.height) {
- tiling->pipe0.width += 1;
- tiling->pipe_count.width =
- DIV_ROUND_UP(tiling->tile_count.width, tiling->pipe0.width);
- } else {
- tiling->pipe0.height += 1;
- tiling->pipe_count.height =
- DIV_ROUND_UP(tiling->tile_count.height, tiling->pipe0.height);
- }
- }
-}
+/* "Normal" cache flushes, that don't require any special handling */
static void
-tu_tiling_config_update_pipes(struct tu_tiling_config *tiling,
- const struct tu_device *dev)
+tu_emit_cache_flush(struct tu_cmd_buffer *cmd_buffer,
+ struct tu_cs *cs)
{
- const uint32_t max_pipe_count = 32; /* A6xx */
- const uint32_t used_pipe_count =
- tiling->pipe_count.width * tiling->pipe_count.height;
- const VkExtent2D last_pipe = {
- .width = (tiling->tile_count.width - 1) % tiling->pipe0.width + 1,
- .height = (tiling->tile_count.height - 1) % tiling->pipe0.height + 1,
- };
-
- assert(used_pipe_count <= max_pipe_count);
- assert(max_pipe_count <= ARRAY_SIZE(tiling->pipe_config));
-
- for (uint32_t y = 0; y < tiling->pipe_count.height; y++) {
- for (uint32_t x = 0; x < tiling->pipe_count.width; x++) {
- const uint32_t pipe_x = tiling->pipe0.width * x;
- const uint32_t pipe_y = tiling->pipe0.height * y;
- const uint32_t pipe_w = (x == tiling->pipe_count.width - 1)
- ? last_pipe.width
- : tiling->pipe0.width;
- const uint32_t pipe_h = (y == tiling->pipe_count.height - 1)
- ? last_pipe.height
- : tiling->pipe0.height;
- const uint32_t n = tiling->pipe_count.width * y + x;
-
- tiling->pipe_config[n] = A6XX_VSC_PIPE_CONFIG_REG_X(pipe_x) |
- A6XX_VSC_PIPE_CONFIG_REG_Y(pipe_y) |
- A6XX_VSC_PIPE_CONFIG_REG_W(pipe_w) |
- A6XX_VSC_PIPE_CONFIG_REG_H(pipe_h);
- tiling->pipe_sizes[n] = CP_SET_BIN_DATA5_0_VSC_SIZE(pipe_w * pipe_h);
- }
- }
-
- memset(tiling->pipe_config + used_pipe_count, 0,
- sizeof(uint32_t) * (max_pipe_count - used_pipe_count));
+ tu6_emit_flushes(cmd_buffer, cs, cmd_buffer->state.cache.flush_bits);
+ cmd_buffer->state.cache.flush_bits = 0;
}
-static void
-tu_tiling_config_get_tile(const struct tu_tiling_config *tiling,
- const struct tu_device *dev,
- uint32_t tx,
- uint32_t ty,
- struct tu_tile *tile)
-{
- /* find the pipe and the slot for tile (tx, ty) */
- const uint32_t px = tx / tiling->pipe0.width;
- const uint32_t py = ty / tiling->pipe0.height;
- const uint32_t sx = tx - tiling->pipe0.width * px;
- const uint32_t sy = ty - tiling->pipe0.height * py;
- /* last pipe has different width */
- const uint32_t pipe_width =
- MIN2(tiling->pipe0.width,
- tiling->tile_count.width - px * tiling->pipe0.width);
-
- assert(tx < tiling->tile_count.width && ty < tiling->tile_count.height);
- assert(px < tiling->pipe_count.width && py < tiling->pipe_count.height);
- assert(sx < tiling->pipe0.width && sy < tiling->pipe0.height);
-
- /* convert to 1D indices */
- tile->pipe = tiling->pipe_count.width * py + px;
- tile->slot = pipe_width * sy + sx;
-
- /* get the blit area for the tile */
- tile->begin = (VkOffset2D) {
- .x = tiling->tile0.offset.x + tiling->tile0.extent.width * tx,
- .y = tiling->tile0.offset.y + tiling->tile0.extent.height * ty,
- };
- tile->end.x =
- (tx == tiling->tile_count.width - 1)
- ? tiling->render_area.offset.x + tiling->render_area.extent.width
- : tile->begin.x + tiling->tile0.extent.width;
- tile->end.y =
- (ty == tiling->tile_count.height - 1)
- ? tiling->render_area.offset.y + tiling->render_area.extent.height
- : tile->begin.y + tiling->tile0.extent.height;
-}
+/* Renderpass cache flushes */
-enum a3xx_msaa_samples
-tu_msaa_samples(uint32_t samples)
+void
+tu_emit_cache_flush_renderpass(struct tu_cmd_buffer *cmd_buffer,
+ struct tu_cs *cs)
{
- switch (samples) {
- case 1:
- return MSAA_ONE;
- case 2:
- return MSAA_TWO;
- case 4:
- return MSAA_FOUR;
- case 8:
- return MSAA_EIGHT;
- default:
- assert(!"invalid sample count");
- return MSAA_ONE;
- }
+ tu6_emit_flushes(cmd_buffer, cs, cmd_buffer->state.renderpass_cache.flush_bits);
+ cmd_buffer->state.renderpass_cache.flush_bits = 0;
}
-static enum a4xx_index_size
-tu6_index_size(VkIndexType type)
-{
- switch (type) {
- case VK_INDEX_TYPE_UINT16:
- return INDEX4_SIZE_16_BIT;
- case VK_INDEX_TYPE_UINT32:
- return INDEX4_SIZE_32_BIT;
- default:
- unreachable("invalid VkIndexType");
- return INDEX4_SIZE_8_BIT;
- }
-}
+/* Cache flushes for things that use the color/depth read/write path (i.e.
+ * blits and draws). This deals with changing CCU state as well as the usual
+ * cache flushing.
+ */
-unsigned
-tu6_emit_event_write(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- enum vgt_event_type event,
- bool need_seqno)
+void
+tu_emit_cache_flush_ccu(struct tu_cmd_buffer *cmd_buffer,
+ struct tu_cs *cs,
+ enum tu_cmd_ccu_state ccu_state)
{
- unsigned seqno = 0;
+ enum tu_cmd_flush_bits flushes = cmd_buffer->state.cache.flush_bits;
- tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, need_seqno ? 4 : 1);
- tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(event));
- if (need_seqno) {
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
- seqno = ++cmd->scratch_seqno;
- tu_cs_emit(cs, seqno);
- }
+ assert(ccu_state != TU_CMD_CCU_UNKNOWN);
- return seqno;
-}
-
-static void
-tu6_emit_cache_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
- tu6_emit_event_write(cmd, cs, 0x31, false);
-}
+ /* Changing CCU state must involve invalidating the CCU. In sysmem mode,
+ * the CCU may also contain data that we haven't flushed out yet, so we
+ * also need to flush. Also, in order to program RB_CCU_CNTL, we need to
+ * emit a WFI as it isn't pipelined.
+ */
+ if (ccu_state != cmd_buffer->state.ccu_state) {
+ if (cmd_buffer->state.ccu_state != TU_CMD_CCU_GMEM) {
+ flushes |=
+ TU_CMD_FLAG_CCU_FLUSH_COLOR |
+ TU_CMD_FLAG_CCU_FLUSH_DEPTH;
+ cmd_buffer->state.cache.pending_flush_bits &= ~(
+ TU_CMD_FLAG_CCU_FLUSH_COLOR |
+ TU_CMD_FLAG_CCU_FLUSH_DEPTH);
+ }
+ flushes |=
+ TU_CMD_FLAG_CCU_INVALIDATE_COLOR |
+ TU_CMD_FLAG_CCU_INVALIDATE_DEPTH |
+ TU_CMD_FLAG_WAIT_FOR_IDLE;
+ cmd_buffer->state.cache.pending_flush_bits &= ~(
+ TU_CMD_FLAG_CCU_INVALIDATE_COLOR |
+ TU_CMD_FLAG_CCU_INVALIDATE_DEPTH |
+ TU_CMD_FLAG_WAIT_FOR_IDLE);
+ }
-static void
-tu6_emit_lrz_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
- tu6_emit_event_write(cmd, cs, LRZ_FLUSH, false);
-}
+ tu6_emit_flushes(cmd_buffer, cs, flushes);
+ cmd_buffer->state.cache.flush_bits = 0;
-static void
-tu6_emit_wfi(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
- if (cmd->wait_for_idle) {
- tu_cs_emit_wfi(cs);
- cmd->wait_for_idle = false;
+ if (ccu_state != cmd_buffer->state.ccu_state) {
+ struct tu_physical_device *phys_dev = cmd_buffer->device->physical_device;
+ tu_cs_emit_regs(cs,
+ A6XX_RB_CCU_CNTL(.offset =
+ ccu_state == TU_CMD_CCU_GMEM ?
+ phys_dev->ccu_offset_gmem :
+ phys_dev->ccu_offset_bypass,
+ .gmem = ccu_state == TU_CMD_CCU_GMEM));
+ cmd_buffer->state.ccu_state = ccu_state;
}
}
-#define tu_image_view_ubwc_pitches(iview) \
- .pitch = tu_image_ubwc_pitch(iview->image, iview->base_mip), \
- .array_pitch = tu_image_ubwc_size(iview->image, iview->base_mip) >> 2
-
static void
tu6_emit_zs(struct tu_cmd_buffer *cmd,
const struct tu_subpass *subpass,
}
const struct tu_image_view *iview = fb->attachments[a].attachment;
- enum a6xx_depth_format fmt = tu6_pipe2depth(iview->vk_format);
+ const struct tu_render_pass_attachment *attachment =
+ &cmd->state.pass->attachments[a];
+ enum a6xx_depth_format fmt = tu6_pipe2depth(attachment->format);
- tu_cs_emit_regs(cs,
- A6XX_RB_DEPTH_BUFFER_INFO(.depth_format = fmt),
- A6XX_RB_DEPTH_BUFFER_PITCH(tu_image_stride(iview->image, iview->base_mip)),
- A6XX_RB_DEPTH_BUFFER_ARRAY_PITCH(iview->image->layout.layer_size),
- A6XX_RB_DEPTH_BUFFER_BASE(tu_image_view_base_ref(iview)),
- A6XX_RB_DEPTH_BUFFER_BASE_GMEM(cmd->state.pass->attachments[a].gmem_offset));
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_BUFFER_INFO, 6);
+ tu_cs_emit(cs, A6XX_RB_DEPTH_BUFFER_INFO(.depth_format = fmt).value);
+ tu_cs_image_ref(cs, iview, 0);
+ tu_cs_emit(cs, attachment->gmem_offset);
tu_cs_emit_regs(cs,
A6XX_GRAS_SU_DEPTH_BUFFER_INFO(.depth_format = fmt));
- tu_cs_emit_regs(cs,
- A6XX_RB_DEPTH_FLAG_BUFFER_BASE(tu_image_view_ubwc_base_ref(iview)),
- A6XX_RB_DEPTH_FLAG_BUFFER_PITCH(tu_image_view_ubwc_pitches(iview)));
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_FLAG_BUFFER_BASE_LO, 3);
+ tu_cs_image_flag_ref(cs, iview, 0);
tu_cs_emit_regs(cs,
A6XX_GRAS_LRZ_BUFFER_BASE(0),
A6XX_GRAS_LRZ_BUFFER_PITCH(0),
A6XX_GRAS_LRZ_FAST_CLEAR_BUFFER_BASE(0));
- tu_cs_emit_regs(cs,
- A6XX_RB_STENCIL_INFO(0));
-
- /* enable zs? */
+ if (attachment->format == VK_FORMAT_S8_UINT) {
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCIL_INFO, 6);
+ tu_cs_emit(cs, A6XX_RB_STENCIL_INFO(.separate_stencil = true).value);
+ tu_cs_image_ref(cs, iview, 0);
+ tu_cs_emit(cs, attachment->gmem_offset);
+ } else {
+ tu_cs_emit_regs(cs,
+ A6XX_RB_STENCIL_INFO(0));
+ }
}
static void
struct tu_cs *cs)
{
const struct tu_framebuffer *fb = cmd->state.framebuffer;
- unsigned char mrt_comp[MAX_RTS] = { 0 };
- unsigned srgb_cntl = 0;
for (uint32_t i = 0; i < subpass->color_count; ++i) {
uint32_t a = subpass->color_attachments[i].attachment;
continue;
const struct tu_image_view *iview = fb->attachments[a].attachment;
- const enum a6xx_tile_mode tile_mode =
- tu6_get_image_tile_mode(iview->image, iview->base_mip);
-
- mrt_comp[i] = 0xf;
-
- if (vk_format_is_srgb(iview->vk_format))
- srgb_cntl |= (1 << i);
-
- const struct tu_native_format format =
- tu6_format_color(iview->vk_format, iview->image->layout.tile_mode);
- tu_cs_emit_regs(cs,
- A6XX_RB_MRT_BUF_INFO(i,
- .color_tile_mode = tile_mode,
- .color_format = format.fmt,
- .color_swap = format.swap),
- A6XX_RB_MRT_PITCH(i, tu_image_stride(iview->image, iview->base_mip)),
- A6XX_RB_MRT_ARRAY_PITCH(i, iview->image->layout.layer_size),
- A6XX_RB_MRT_BASE(i, tu_image_view_base_ref(iview)),
- A6XX_RB_MRT_BASE_GMEM(i, cmd->state.pass->attachments[a].gmem_offset));
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_BUF_INFO(i), 6);
+ tu_cs_emit(cs, iview->RB_MRT_BUF_INFO);
+ tu_cs_image_ref(cs, iview, 0);
+ tu_cs_emit(cs, cmd->state.pass->attachments[a].gmem_offset);
tu_cs_emit_regs(cs,
- A6XX_SP_FS_MRT_REG(i,
- .color_format = format.fmt,
- .color_sint = vk_format_is_sint(iview->vk_format),
- .color_uint = vk_format_is_uint(iview->vk_format)));
+ A6XX_SP_FS_MRT_REG(i, .dword = iview->SP_FS_MRT_REG));
- tu_cs_emit_regs(cs,
- A6XX_RB_MRT_FLAG_BUFFER_ADDR(i, tu_image_view_ubwc_base_ref(iview)),
- A6XX_RB_MRT_FLAG_BUFFER_PITCH(i, tu_image_view_ubwc_pitches(iview)));
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_FLAG_BUFFER_ADDR_LO(i), 3);
+ tu_cs_image_flag_ref(cs, iview, 0);
}
tu_cs_emit_regs(cs,
- A6XX_RB_SRGB_CNTL(.dword = srgb_cntl));
-
- tu_cs_emit_regs(cs,
- A6XX_SP_SRGB_CNTL(.dword = srgb_cntl));
-
+ A6XX_RB_SRGB_CNTL(.dword = subpass->srgb_cntl));
tu_cs_emit_regs(cs,
- A6XX_RB_RENDER_COMPONENTS(
- .rt0 = mrt_comp[0],
- .rt1 = mrt_comp[1],
- .rt2 = mrt_comp[2],
- .rt3 = mrt_comp[3],
- .rt4 = mrt_comp[4],
- .rt5 = mrt_comp[5],
- .rt6 = mrt_comp[6],
- .rt7 = mrt_comp[7]));
+ A6XX_SP_SRGB_CNTL(.dword = subpass->srgb_cntl));
- tu_cs_emit_regs(cs,
- A6XX_SP_FS_RENDER_COMPONENTS(
- .rt0 = mrt_comp[0],
- .rt1 = mrt_comp[1],
- .rt2 = mrt_comp[2],
- .rt3 = mrt_comp[3],
- .rt4 = mrt_comp[4],
- .rt5 = mrt_comp[5],
- .rt6 = mrt_comp[6],
- .rt7 = mrt_comp[7]));
-
- // XXX: We probably can't hardcode LAYER_CNTL_TYPE.
- tu_cs_emit_regs(cs,
- A6XX_GRAS_LAYER_CNTL(.layered = fb->layers > 1,
- .type = LAYER_2D_ARRAY));
+ tu_cs_emit_regs(cs, A6XX_GRAS_MAX_LAYER_INDEX(fb->layers - 1));
}
-static void
-tu6_emit_msaa(struct tu_cmd_buffer *cmd,
- const struct tu_subpass *subpass,
- struct tu_cs *cs)
+void
+tu6_emit_msaa(struct tu_cs *cs, VkSampleCountFlagBits vk_samples)
{
- const enum a3xx_msaa_samples samples = tu_msaa_samples(subpass->samples);
+ const enum a3xx_msaa_samples samples = tu_msaa_samples(vk_samples);
bool msaa_disable = samples == MSAA_ONE;
tu_cs_emit_regs(cs,
continue;
const struct tu_image_view *iview = fb->attachments[a].attachment;
- if (iview->image->layout.ubwc_layer_size != 0)
+ if (iview->ubwc_enabled)
mrts_ubwc_enable |= 1 << i;
}
const uint32_t a = subpass->depth_stencil_attachment.attachment;
if (a != VK_ATTACHMENT_UNUSED) {
const struct tu_image_view *iview = fb->attachments[a].attachment;
- if (iview->image->layout.ubwc_layer_size != 0)
+ if (iview->ubwc_enabled)
cntl |= A6XX_RB_RENDER_CNTL_FLAG_DEPTH;
}
static void
tu6_emit_blit_scissor(struct tu_cmd_buffer *cmd, struct tu_cs *cs, bool align)
{
- const VkRect2D *render_area = &cmd->state.tiling_config.render_area;
+
+ const VkRect2D *render_area = &cmd->state.render_area;
+
+ /* Avoid assertion fails with an empty render area at (0, 0) where the
+ * subtraction below wraps around. Empty render areas should be forced to
+ * the sysmem path by use_sysmem_rendering(). It's not even clear whether
+ * an empty scissor here works, and the blob seems to force sysmem too as
+ * it sets something wrong (non-empty) for the scissor.
+ */
+ if (render_area->extent.width == 0 ||
+ render_area->extent.height == 0)
+ return;
+
uint32_t x1 = render_area->offset.x;
uint32_t y1 = render_area->offset.y;
uint32_t x2 = x1 + render_area->extent.width - 1;
uint32_t y2 = y1 + render_area->extent.height - 1;
- /* TODO: alignment requirement seems to be less than tile_align_w/h */
if (align) {
- x1 = x1 & ~cmd->device->physical_device->tile_align_w;
- y1 = y1 & ~cmd->device->physical_device->tile_align_h;
- x2 = ALIGN_POT(x2 + 1, cmd->device->physical_device->tile_align_w) - 1;
- y2 = ALIGN_POT(y2 + 1, cmd->device->physical_device->tile_align_h) - 1;
+ x1 = x1 & ~(GMEM_ALIGN_W - 1);
+ y1 = y1 & ~(GMEM_ALIGN_H - 1);
+ x2 = ALIGN_POT(x2 + 1, GMEM_ALIGN_W) - 1;
+ y2 = ALIGN_POT(y2 + 1, GMEM_ALIGN_H) - 1;
}
tu_cs_emit_regs(cs,
A6XX_RB_BLIT_SCISSOR_BR(.x = x2, .y = y2));
}
-static void
-tu6_emit_blit_info(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_image_view *iview,
- uint32_t gmem_offset,
- bool resolve)
-{
- tu_cs_emit_regs(cs,
- A6XX_RB_BLIT_INFO(.unk0 = !resolve, .gmem = !resolve));
-
- const struct tu_native_format format =
- tu6_format_color(iview->vk_format, iview->image->layout.tile_mode);
-
- enum a6xx_tile_mode tile_mode =
- tu6_get_image_tile_mode(iview->image, iview->base_mip);
- tu_cs_emit_regs(cs,
- A6XX_RB_BLIT_DST_INFO(
- .tile_mode = tile_mode,
- .samples = tu_msaa_samples(iview->image->samples),
- .color_format = format.fmt,
- .color_swap = format.swap,
- .flags = iview->image->layout.ubwc_layer_size != 0),
- A6XX_RB_BLIT_DST(tu_image_view_base_ref(iview)),
- A6XX_RB_BLIT_DST_PITCH(tu_image_stride(iview->image, iview->base_mip)),
- A6XX_RB_BLIT_DST_ARRAY_PITCH(iview->image->layout.layer_size));
-
- if (iview->image->layout.ubwc_layer_size) {
- tu_cs_emit_regs(cs,
- A6XX_RB_BLIT_FLAG_DST(tu_image_view_ubwc_base_ref(iview)),
- A6XX_RB_BLIT_FLAG_DST_PITCH(tu_image_view_ubwc_pitches(iview)));
- }
-
- tu_cs_emit_regs(cs,
- A6XX_RB_BLIT_BASE_GMEM(gmem_offset));
-}
-
-static void
-tu6_emit_blit(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
- tu6_emit_event_write(cmd, cs, BLIT, false);
-}
-
-static void
-tu6_emit_window_scissor(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
+void
+tu6_emit_window_scissor(struct tu_cs *cs,
uint32_t x1,
uint32_t y1,
uint32_t x2,
A6XX_GRAS_SC_WINDOW_SCISSOR_BR(.x = x2, .y = y2));
tu_cs_emit_regs(cs,
- A6XX_GRAS_RESOLVE_CNTL_1(.x = x1, .y = y1),
- A6XX_GRAS_RESOLVE_CNTL_2(.x = x2, .y = y2));
+ A6XX_GRAS_2D_RESOLVE_CNTL_1(.x = x1, .y = y1),
+ A6XX_GRAS_2D_RESOLVE_CNTL_2(.x = x2, .y = y2));
}
-static void
-tu6_emit_window_offset(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- uint32_t x1,
- uint32_t y1)
+void
+tu6_emit_window_offset(struct tu_cs *cs, uint32_t x1, uint32_t y1)
{
tu_cs_emit_regs(cs,
A6XX_RB_WINDOW_OFFSET(.x = x1, .y = y1));
A6XX_SP_TP_WINDOW_OFFSET(.x = x1, .y = y1));
}
+static void
+tu_cs_emit_draw_state(struct tu_cs *cs, uint32_t id, struct tu_draw_state state)
+{
+ uint32_t enable_mask;
+ switch (id) {
+ case TU_DRAW_STATE_PROGRAM:
+ case TU_DRAW_STATE_VI:
+ case TU_DRAW_STATE_FS_CONST:
+ /* The blob seems to not enable this (DESC_SETS_LOAD) for binning, even
+ * when resources would actually be used in the binning shader.
+ * Presumably the overhead of prefetching the resources isn't
+ * worth it.
+ */
+ case TU_DRAW_STATE_DESC_SETS_LOAD:
+ enable_mask = CP_SET_DRAW_STATE__0_GMEM |
+ CP_SET_DRAW_STATE__0_SYSMEM;
+ break;
+ case TU_DRAW_STATE_PROGRAM_BINNING:
+ case TU_DRAW_STATE_VI_BINNING:
+ enable_mask = CP_SET_DRAW_STATE__0_BINNING;
+ break;
+ case TU_DRAW_STATE_INPUT_ATTACHMENTS_GMEM:
+ enable_mask = CP_SET_DRAW_STATE__0_GMEM;
+ break;
+ case TU_DRAW_STATE_INPUT_ATTACHMENTS_SYSMEM:
+ enable_mask = CP_SET_DRAW_STATE__0_SYSMEM;
+ break;
+ default:
+ enable_mask = CP_SET_DRAW_STATE__0_GMEM |
+ CP_SET_DRAW_STATE__0_SYSMEM |
+ CP_SET_DRAW_STATE__0_BINNING;
+ break;
+ }
+
+ /* We need to reload the descriptors every time the descriptor sets
+ * change. However, the commands we send only depend on the pipeline
+ * because the whole point is to cache descriptors which are used by the
+ * pipeline. There's a problem here, in that the firmware has an
+ * "optimization" which skips executing groups that are set to the same
+ * value as the last draw. This means that if the descriptor sets change
+ * but not the pipeline, we'd try to re-execute the same buffer which
+ * the firmware would ignore and we wouldn't pre-load the new
+ * descriptors. Set the DIRTY bit to avoid this optimization
+ */
+ if (id == TU_DRAW_STATE_DESC_SETS_LOAD)
+ enable_mask |= CP_SET_DRAW_STATE__0_DIRTY;
+
+ tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(state.size) |
+ enable_mask |
+ CP_SET_DRAW_STATE__0_GROUP_ID(id) |
+ COND(!state.size, CP_SET_DRAW_STATE__0_DISABLE));
+ tu_cs_emit_qw(cs, state.iova);
+}
+
static bool
use_hw_binning(struct tu_cmd_buffer *cmd)
{
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
+
+ /* XFB commands are emitted for BINNING || SYSMEM, which makes it incompatible
+ * with non-hw binning GMEM rendering. this is required because some of the
+ * XFB commands need to only be executed once
+ */
+ if (cmd->state.xfb_used)
+ return true;
if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_NOBIN))
return false;
if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_FORCEBIN))
return true;
- return (tiling->tile_count.width * tiling->tile_count.height) > 2;
+ return (fb->tile_count.width * fb->tile_count.height) > 2;
}
static bool
if (!cmd->state.pass->gmem_pixels)
return true;
- return cmd->state.tiling_config.force_sysmem;
+ if (cmd->state.framebuffer->layers > 1)
+ return true;
+
+ /* Use sysmem for empty render areas */
+ if (cmd->state.render_area.extent.width == 0 ||
+ cmd->state.render_area.extent.height == 0)
+ return true;
+
+ if (cmd->has_tess)
+ return true;
+
+ return false;
}
static void
tu6_emit_tile_select(struct tu_cmd_buffer *cmd,
struct tu_cs *cs,
- const struct tu_tile *tile)
+ uint32_t tx, uint32_t ty, uint32_t pipe, uint32_t slot)
{
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
+
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_YIELD));
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_GMEM));
- const uint32_t x1 = tile->begin.x;
- const uint32_t y1 = tile->begin.y;
- const uint32_t x2 = tile->end.x - 1;
- const uint32_t y2 = tile->end.y - 1;
- tu6_emit_window_scissor(cmd, cs, x1, y1, x2, y2);
- tu6_emit_window_offset(cmd, cs, x1, y1);
+ const uint32_t x1 = fb->tile0.width * tx;
+ const uint32_t y1 = fb->tile0.height * ty;
+ const uint32_t x2 = x1 + fb->tile0.width - 1;
+ const uint32_t y2 = y1 + fb->tile0.height - 1;
+ tu6_emit_window_scissor(cs, x1, y1, x2, y2);
+ tu6_emit_window_offset(cs, x1, y1);
- tu_cs_emit_regs(cs,
- A6XX_VPC_SO_OVERRIDE(.so_disable = false));
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(false));
if (use_hw_binning(cmd)) {
tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
tu_cs_emit(cs, 0x0);
- tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
- tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
- A6XX_CP_REG_TEST_0_BIT(0) |
- A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
+ tu_cs_emit_pkt7(cs, CP_SET_BIN_DATA5_OFFSET, 4);
+ tu_cs_emit(cs, fb->pipe_sizes[pipe] |
+ CP_SET_BIN_DATA5_0_VSC_N(slot));
+ tu_cs_emit(cs, pipe * cmd->vsc_draw_strm_pitch);
+ tu_cs_emit(cs, pipe * 4);
+ tu_cs_emit(cs, pipe * cmd->vsc_prim_strm_pitch);
- tu_cs_reserve(cs, 3 + 11);
- tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
- tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
- tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(11));
-
- /* if (no overflow) */ {
- tu_cs_emit_pkt7(cs, CP_SET_BIN_DATA5, 7);
- tu_cs_emit(cs, cmd->state.tiling_config.pipe_sizes[tile->pipe] |
- CP_SET_BIN_DATA5_0_VSC_N(tile->slot));
- tu_cs_emit_qw(cs, cmd->vsc_data.iova + tile->pipe * cmd->vsc_data_pitch);
- tu_cs_emit_qw(cs, cmd->vsc_data.iova + (tile->pipe * 4) + (32 * cmd->vsc_data_pitch));
- tu_cs_emit_qw(cs, cmd->vsc_data2.iova + (tile->pipe * cmd->vsc_data2_pitch));
-
- tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
- tu_cs_emit(cs, 0x0);
-
- /* use a NOP packet to skip over the 'else' side: */
- tu_cs_emit_pkt7(cs, CP_NOP, 2);
- } /* else */ {
- tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
- tu_cs_emit(cs, 0x1);
- }
+ tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
+ tu_cs_emit(cs, 0x0);
tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
tu_cs_emit(cs, 0x0);
-
- tu_cs_emit_regs(cs,
- A6XX_RB_UNKNOWN_8804(0));
-
- tu_cs_emit_regs(cs,
- A6XX_SP_TP_UNKNOWN_B304(0));
-
- tu_cs_emit_regs(cs,
- A6XX_GRAS_UNKNOWN_80A4(0));
} else {
tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
tu_cs_emit(cs, 0x1);
}
}
-static void
-tu6_emit_load_attachment(struct tu_cmd_buffer *cmd, struct tu_cs *cs, uint32_t a)
-{
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
- const struct tu_framebuffer *fb = cmd->state.framebuffer;
- const struct tu_image_view *iview = fb->attachments[a].attachment;
- const struct tu_render_pass_attachment *attachment =
- &cmd->state.pass->attachments[a];
-
- if (attachment->gmem_offset < 0)
- return;
-
- const uint32_t x1 = tiling->render_area.offset.x;
- const uint32_t y1 = tiling->render_area.offset.y;
- const uint32_t x2 = x1 + tiling->render_area.extent.width;
- const uint32_t y2 = y1 + tiling->render_area.extent.height;
- const uint32_t tile_x2 =
- tiling->tile0.offset.x + tiling->tile0.extent.width * tiling->tile_count.width;
- const uint32_t tile_y2 =
- tiling->tile0.offset.y + tiling->tile0.extent.height * tiling->tile_count.height;
- bool need_load =
- x1 != tiling->tile0.offset.x || x2 != MIN2(fb->width, tile_x2) ||
- y1 != tiling->tile0.offset.y || y2 != MIN2(fb->height, tile_y2);
-
- if (need_load)
- tu_finishme("improve handling of unaligned render area");
-
- if (attachment->load_op == VK_ATTACHMENT_LOAD_OP_LOAD)
- need_load = true;
-
- if (vk_format_has_stencil(iview->vk_format) &&
- attachment->stencil_load_op == VK_ATTACHMENT_LOAD_OP_LOAD)
- need_load = true;
-
- if (need_load) {
- tu6_emit_blit_info(cmd, cs, iview, attachment->gmem_offset, false);
- tu6_emit_blit(cmd, cs);
- }
-}
-
-static void
-tu6_emit_clear_attachment(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
- uint32_t a,
- const VkRenderPassBeginInfo *info)
-{
- const struct tu_framebuffer *fb = cmd->state.framebuffer;
- const struct tu_image_view *iview = fb->attachments[a].attachment;
- const struct tu_render_pass_attachment *attachment =
- &cmd->state.pass->attachments[a];
- unsigned clear_mask = 0;
-
- /* note: this means it isn't used by any subpass and shouldn't be cleared anyway */
- if (attachment->gmem_offset < 0)
- return;
-
- if (attachment->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR)
- clear_mask = 0xf;
-
- if (vk_format_has_stencil(iview->vk_format)) {
- clear_mask &= 0x1;
- if (attachment->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR)
- clear_mask |= 0x2;
- }
- if (!clear_mask)
- return;
-
- tu_clear_gmem_attachment(cmd, cs, a, clear_mask,
- &info->pClearValues[a]);
-}
-
-static void
-tu6_emit_predicated_blit(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- uint32_t a,
- uint32_t gmem_a,
- bool resolve)
-{
- tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_GMEM);
-
- tu6_emit_blit_info(cmd, cs,
- cmd->state.framebuffer->attachments[a].attachment,
- cmd->state.pass->attachments[gmem_a].gmem_offset, resolve);
- tu6_emit_blit(cmd, cs);
-
- tu_cond_exec_end(cs);
-}
-
static void
tu6_emit_sysmem_resolve(struct tu_cmd_buffer *cmd,
struct tu_cs *cs,
uint32_t gmem_a)
{
const struct tu_framebuffer *fb = cmd->state.framebuffer;
- const struct tu_image_view *dst = fb->attachments[a].attachment;
- const struct tu_image_view *src = fb->attachments[gmem_a].attachment;
-
- tu_blit(cmd, cs, &(struct tu_blit) {
- .dst = sysmem_attachment_surf(dst, dst->base_layer,
- &cmd->state.tiling_config.render_area),
- .src = sysmem_attachment_surf(src, src->base_layer,
- &cmd->state.tiling_config.render_area),
- .layers = fb->layers,
- });
-}
-
+ struct tu_image_view *dst = fb->attachments[a].attachment;
+ struct tu_image_view *src = fb->attachments[gmem_a].attachment;
-/* Emit a MSAA resolve operation, with both gmem and sysmem paths. */
-static void tu6_emit_resolve(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- uint32_t a,
- uint32_t gmem_a)
-{
- if (cmd->state.pass->attachments[a].store_op == VK_ATTACHMENT_STORE_OP_DONT_CARE)
- return;
-
- tu6_emit_predicated_blit(cmd, cs, a, gmem_a, true);
-
- tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
- tu6_emit_sysmem_resolve(cmd, cs, a, gmem_a);
- tu_cond_exec_end(cs);
+ tu_resolve_sysmem(cmd, cs, src, dst, fb->layers, &cmd->state.render_area);
}
static void
-tu6_emit_store_attachment(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- uint32_t a,
- uint32_t gmem_a)
+tu6_emit_sysmem_resolves(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs,
+ const struct tu_subpass *subpass)
{
- if (cmd->state.pass->attachments[a].store_op == VK_ATTACHMENT_STORE_OP_DONT_CARE)
- return;
+ if (subpass->resolve_attachments) {
+ /* From the documentation for vkCmdNextSubpass, section 7.4 "Render Pass
+ * Commands":
+ *
+ * End-of-subpass multisample resolves are treated as color
+ * attachment writes for the purposes of synchronization. That is,
+ * they are considered to execute in the
+ * VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT pipeline stage and
+ * their writes are synchronized with
+ * VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT. Synchronization between
+ * rendering within a subpass and any resolve operations at the end
+ * of the subpass occurs automatically, without need for explicit
+ * dependencies or pipeline barriers. However, if the resolve
+ * attachment is also used in a different subpass, an explicit
+ * dependency is needed.
+ *
+ * We use the CP_BLIT path for sysmem resolves, which is really a
+ * transfer command, so we have to manually flush similar to the gmem
+ * resolve case. However, a flush afterwards isn't needed because of the
+ * last sentence and the fact that we're in sysmem mode.
+ */
+ tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS);
+ tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE);
+
+ /* Wait for the flushes to land before using the 2D engine */
+ tu_cs_emit_wfi(cs);
+
+ for (unsigned i = 0; i < subpass->color_count; i++) {
+ uint32_t a = subpass->resolve_attachments[i].attachment;
+ if (a == VK_ATTACHMENT_UNUSED)
+ continue;
- tu6_emit_blit_info(cmd, cs,
- cmd->state.framebuffer->attachments[a].attachment,
- cmd->state.pass->attachments[gmem_a].gmem_offset, true);
- tu6_emit_blit(cmd, cs);
+ tu6_emit_sysmem_resolve(cmd, cs, a,
+ subpass->color_attachments[i].attachment);
+ }
+ }
}
static void
for (uint32_t a = 0; a < pass->attachment_count; ++a) {
if (pass->attachments[a].gmem_offset >= 0)
- tu6_emit_store_attachment(cmd, cs, a, a);
+ tu_store_gmem_attachment(cmd, cs, a, a);
}
if (subpass->resolve_attachments) {
for (unsigned i = 0; i < subpass->color_count; i++) {
uint32_t a = subpass->resolve_attachments[i].attachment;
if (a != VK_ATTACHMENT_UNUSED)
- tu6_emit_store_attachment(cmd, cs, a,
- subpass->color_attachments[i].attachment);
+ tu_store_gmem_attachment(cmd, cs, a,
+ subpass->color_attachments[i].attachment);
}
}
}
-static void
-tu6_emit_restart_index(struct tu_cs *cs, uint32_t restart_index)
-{
- tu_cs_emit_regs(cs,
- A6XX_PC_RESTART_INDEX(restart_index));
-}
-
static void
tu6_init_hw(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
- tu6_emit_cache_flush(cmd, cs);
+ struct tu_device *dev = cmd->device;
+ const struct tu_physical_device *phys_dev = dev->physical_device;
+
+ tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE);
+
+ tu_cs_emit_regs(cs, A6XX_HLSQ_INVALIDATE_CMD(
+ .vs_state = true,
+ .hs_state = true,
+ .ds_state = true,
+ .gs_state = true,
+ .fs_state = true,
+ .cs_state = true,
+ .gfx_ibo = true,
+ .cs_ibo = true,
+ .gfx_shared_const = true,
+ .cs_shared_const = true,
+ .gfx_bindless = 0x1f,
+ .cs_bindless = 0x1f));
+
+ tu_cs_emit_wfi(cs);
- tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UPDATE_CNTL, 0xfffff);
+ cmd->state.cache.pending_flush_bits &=
+ ~(TU_CMD_FLAG_WAIT_FOR_IDLE | TU_CMD_FLAG_CACHE_INVALIDATE);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_CCU_CNTL, 0x10000000);
+ tu_cs_emit_regs(cs,
+ A6XX_RB_CCU_CNTL(.offset = phys_dev->ccu_offset_bypass));
+ cmd->state.ccu_state = TU_CMD_CCU_SYSMEM;
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8E04, 0x00100000);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE04, 0x8);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE00, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AE03, 0x00000410);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_IBO_COUNT, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B182, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_UNKNOWN_BB11, 0);
+ tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_SHARED_CONSTS, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_UCHE_UNKNOWN_0E12, 0x3200000);
tu_cs_emit_write_reg(cs, REG_A6XX_UCHE_CLIENT_PF, 4);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8E01, 0x0);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_A982, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_A9A8, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_AB00, 0x5);
- tu_cs_emit_write_reg(cs, REG_A6XX_VPC_GS_SIV_CNTL, 0x0000ffff);
+ /* TODO: set A6XX_VFD_ADD_OFFSET_INSTANCE and fix ir3 to avoid adding base instance */
tu_cs_emit_write_reg(cs, REG_A6XX_VFD_ADD_OFFSET, A6XX_VFD_ADD_OFFSET_VERTEX);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8811, 0x00000010);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_MODE_CNTL, 0x1f);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881E, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_88F0, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9101, 0xffff00);
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9107, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9236, 1);
+ tu_cs_emit_regs(cs, A6XX_VPC_POINT_COORD_INVERT(false));
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9300, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_VPC_SO_OVERRIDE,
- A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(true));
- tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9801, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9806, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9980, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9990, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_PRIMITIVE_CNTL_6, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9B07, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_UNKNOWN_B183, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8099, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_809B, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A0, 2);
tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80AF, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9210, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9211, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9602, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9981, 0x3);
tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9E72, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_VPC_UNKNOWN_9108, 0x3);
- tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_UNKNOWN_B304, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_SP_TP_UNKNOWN_B309, 0x000000a2);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8804, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A4, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A5, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_80A6, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8805, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8806, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8878, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8879, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_HLSQ_CONTROL_5_REG, 0xfc);
tu_cs_emit_write_reg(cs, REG_A6XX_VFD_MODE_CNTL, 0x00000000);
tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
- /* Set not to use streamout by default, */
- tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
- tu_cs_emit(cs, 0);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
- tu_cs_emit(cs, 0);
-
tu_cs_emit_regs(cs,
A6XX_SP_HS_CTRL_REG0(0));
A6XX_RB_LRZ_CNTL(0));
tu_cs_emit_regs(cs,
- A6XX_SP_TP_BORDER_COLOR_BASE_ADDR(.bo = &cmd->device->border_color));
+ A6XX_SP_TP_BORDER_COLOR_BASE_ADDR(.bo = &dev->global_bo,
+ .bo_offset = gb_offset(border_color)));
tu_cs_emit_regs(cs,
- A6XX_SP_PS_TP_BORDER_COLOR_BASE_ADDR(.bo = &cmd->device->border_color));
+ A6XX_SP_PS_TP_BORDER_COLOR_BASE_ADDR(.bo = &dev->global_bo,
+ .bo_offset = gb_offset(border_color)));
- tu_cs_sanity_check(cs);
-}
+ /* VSC buffers:
+ * use vsc pitches from the largest values used so far with this device
+ * if there hasn't been overflow, there will already be a scratch bo
+ * allocated for these sizes
+ *
+ * if overflow is detected, the stream size is increased by 2x
+ */
+ mtx_lock(&dev->vsc_pitch_mtx);
-static void
-tu6_cache_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
- unsigned seqno;
+ struct tu6_global *global = dev->global_bo.map;
- seqno = tu6_emit_event_write(cmd, cs, RB_DONE_TS, true);
+ uint32_t vsc_draw_overflow = global->vsc_draw_overflow;
+ uint32_t vsc_prim_overflow = global->vsc_prim_overflow;
- tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
- tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
- CP_WAIT_REG_MEM_0_POLL_MEMORY);
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
- tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(seqno));
- tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0));
- tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(16));
+ if (vsc_draw_overflow >= dev->vsc_draw_strm_pitch)
+ dev->vsc_draw_strm_pitch = (dev->vsc_draw_strm_pitch - VSC_PAD) * 2 + VSC_PAD;
- seqno = tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS, true);
+ if (vsc_prim_overflow >= dev->vsc_prim_strm_pitch)
+ dev->vsc_prim_strm_pitch = (dev->vsc_prim_strm_pitch - VSC_PAD) * 2 + VSC_PAD;
- tu_cs_emit_pkt7(cs, CP_WAIT_MEM_GTE, 4);
- tu_cs_emit(cs, CP_WAIT_MEM_GTE_0_RESERVED(0));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova);
- tu_cs_emit(cs, CP_WAIT_MEM_GTE_3_REF(seqno));
-}
+ cmd->vsc_prim_strm_pitch = dev->vsc_prim_strm_pitch;
+ cmd->vsc_draw_strm_pitch = dev->vsc_draw_strm_pitch;
-static void
-update_vsc_pipe(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
-{
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+ mtx_unlock(&dev->vsc_pitch_mtx);
+
+ struct tu_bo *vsc_bo;
+ uint32_t size0 = cmd->vsc_prim_strm_pitch * MAX_VSC_PIPES +
+ cmd->vsc_draw_strm_pitch * MAX_VSC_PIPES;
+
+ tu_get_scratch_bo(dev, size0 + MAX_VSC_PIPES * 4, &vsc_bo);
+
+ tu_cs_emit_regs(cs,
+ A6XX_VSC_DRAW_STRM_SIZE_ADDRESS(.bo = vsc_bo, .bo_offset = size0));
+ tu_cs_emit_regs(cs,
+ A6XX_VSC_PRIM_STRM_ADDRESS(.bo = vsc_bo));
+ tu_cs_emit_regs(cs,
+ A6XX_VSC_DRAW_STRM_ADDRESS(.bo = vsc_bo,
+ .bo_offset = cmd->vsc_prim_strm_pitch * MAX_VSC_PIPES));
+
+ tu_bo_list_add(&cmd->bo_list, vsc_bo, MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
+
+ tu_cs_sanity_check(cs);
+}
+
+static void
+update_vsc_pipe(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
tu_cs_emit_regs(cs,
- A6XX_VSC_BIN_SIZE(.width = tiling->tile0.extent.width,
- .height = tiling->tile0.extent.height),
- A6XX_VSC_SIZE_ADDRESS(.bo = &cmd->vsc_data,
- .bo_offset = 32 * cmd->vsc_data_pitch));
+ A6XX_VSC_BIN_SIZE(.width = fb->tile0.width,
+ .height = fb->tile0.height));
tu_cs_emit_regs(cs,
- A6XX_VSC_BIN_COUNT(.nx = tiling->tile_count.width,
- .ny = tiling->tile_count.height));
+ A6XX_VSC_BIN_COUNT(.nx = fb->tile_count.width,
+ .ny = fb->tile_count.height));
tu_cs_emit_pkt4(cs, REG_A6XX_VSC_PIPE_CONFIG_REG(0), 32);
- for (unsigned i = 0; i < 32; i++)
- tu_cs_emit(cs, tiling->pipe_config[i]);
+ tu_cs_emit_array(cs, fb->pipe_config, 32);
tu_cs_emit_regs(cs,
- A6XX_VSC_PIPE_DATA2_ADDRESS(.bo = &cmd->vsc_data2),
- A6XX_VSC_PIPE_DATA2_PITCH(cmd->vsc_data2_pitch),
- A6XX_VSC_PIPE_DATA2_ARRAY_PITCH(cmd->vsc_data2.size));
+ A6XX_VSC_PRIM_STRM_PITCH(cmd->vsc_prim_strm_pitch),
+ A6XX_VSC_PRIM_STRM_LIMIT(cmd->vsc_prim_strm_pitch - VSC_PAD));
tu_cs_emit_regs(cs,
- A6XX_VSC_PIPE_DATA_ADDRESS(.bo = &cmd->vsc_data),
- A6XX_VSC_PIPE_DATA_PITCH(cmd->vsc_data_pitch),
- A6XX_VSC_PIPE_DATA_ARRAY_PITCH(cmd->vsc_data.size));
+ A6XX_VSC_DRAW_STRM_PITCH(cmd->vsc_draw_strm_pitch),
+ A6XX_VSC_DRAW_STRM_LIMIT(cmd->vsc_draw_strm_pitch - VSC_PAD));
}
static void
emit_vsc_overflow_test(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
const uint32_t used_pipe_count =
- tiling->pipe_count.width * tiling->pipe_count.height;
+ fb->pipe_count.width * fb->pipe_count.height;
- /* Clear vsc_scratch: */
- tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_scratch));
- tu_cs_emit(cs, 0x0);
-
- /* Check for overflow, write vsc_scratch if detected: */
for (int i = 0; i < used_pipe_count; i++) {
tu_cs_emit_pkt7(cs, CP_COND_WRITE5, 8);
tu_cs_emit(cs, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
CP_COND_WRITE5_0_WRITE_MEMORY);
- tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_SIZE_REG(i)));
+ tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_DRAW_STRM_SIZE_REG(i)));
tu_cs_emit(cs, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
- tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_data_pitch));
+ tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_draw_strm_pitch - VSC_PAD));
tu_cs_emit(cs, CP_COND_WRITE5_4_MASK(~0));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_scratch));
- tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(1 + cmd->vsc_data_pitch));
+ tu_cs_emit_qw(cs, global_iova(cmd, vsc_draw_overflow));
+ tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(cmd->vsc_draw_strm_pitch));
tu_cs_emit_pkt7(cs, CP_COND_WRITE5, 8);
tu_cs_emit(cs, CP_COND_WRITE5_0_FUNCTION(WRITE_GE) |
CP_COND_WRITE5_0_WRITE_MEMORY);
- tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_SIZE2_REG(i)));
+ tu_cs_emit(cs, CP_COND_WRITE5_1_POLL_ADDR_LO(REG_A6XX_VSC_PRIM_STRM_SIZE_REG(i)));
tu_cs_emit(cs, CP_COND_WRITE5_2_POLL_ADDR_HI(0));
- tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_data2_pitch));
+ tu_cs_emit(cs, CP_COND_WRITE5_3_REF(cmd->vsc_prim_strm_pitch - VSC_PAD));
tu_cs_emit(cs, CP_COND_WRITE5_4_MASK(~0));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_scratch));
- tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(3 + cmd->vsc_data2_pitch));
+ tu_cs_emit_qw(cs, global_iova(cmd, vsc_prim_overflow));
+ tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(cmd->vsc_prim_strm_pitch));
}
tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
-
- tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
-
- tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
- tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(OVERFLOW_FLAG_REG) |
- CP_MEM_TO_REG_0_CNT(1 - 1));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_scratch));
-
- /*
- * This is a bit awkward, we really want a way to invert the
- * CP_REG_TEST/CP_COND_REG_EXEC logic, so that we can conditionally
- * execute cmds to use hwbinning when a bit is *not* set. This
- * dance is to invert OVERFLOW_FLAG_REG
- *
- * A CP_NOP packet is used to skip executing the 'else' clause
- * if (b0 set)..
- */
-
- /* b0 will be set if VSC_DATA or VSC_DATA2 overflow: */
- tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
- tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
- A6XX_CP_REG_TEST_0_BIT(0) |
- A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
-
- tu_cs_reserve(cs, 3 + 7);
- tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
- tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
- tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(7));
-
- /* if (b0 set) */ {
- /*
- * On overflow, mirror the value to control->vsc_overflow
- * which CPU is checking to detect overflow (see
- * check_vsc_overflow())
- */
- tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
- tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(OVERFLOW_FLAG_REG) |
- CP_REG_TO_MEM_0_CNT(0));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova + ctrl_offset(vsc_overflow));
-
- tu_cs_emit_pkt4(cs, OVERFLOW_FLAG_REG, 1);
- tu_cs_emit(cs, 0x0);
-
- tu_cs_emit_pkt7(cs, CP_NOP, 2); /* skip 'else' when 'if' is taken */
- } /* else */ {
- tu_cs_emit_pkt4(cs, OVERFLOW_FLAG_REG, 1);
- tu_cs_emit(cs, 0x1);
- }
}
static void
tu6_emit_binning_pass(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
struct tu_physical_device *phys_dev = cmd->device->physical_device;
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
-
- uint32_t x1 = tiling->tile0.offset.x;
- uint32_t y1 = tiling->tile0.offset.y;
- uint32_t x2 = tiling->render_area.offset.x + tiling->render_area.extent.width - 1;
- uint32_t y2 = tiling->render_area.offset.y + tiling->render_area.extent.height - 1;
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
- tu6_emit_window_scissor(cmd, cs, x1, y1, x2, y2);
+ tu6_emit_window_scissor(cs, 0, 0, fb->width - 1, fb->height - 1);
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_BINNING));
tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
tu_cs_emit(cs, UNK_2D);
- tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
- tu6_cache_flush(cmd, cs);
+ /* This flush is probably required because the VSC, which produces the
+ * visibility stream, is a client of UCHE, whereas the CP needs to read the
+ * visibility stream (without caching) to do draw skipping. The
+ * WFI+WAIT_FOR_ME combination guarantees that the binning commands
+ * submitted are finished before reading the VSC regs (in
+ * emit_vsc_overflow_test) or the VSC_DATA buffer directly (implicitly as
+ * part of draws).
+ */
+ tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS);
tu_cs_emit_wfi(cs);
tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
tu_cs_emit(cs, 0x0);
+}
- tu_cs_emit_wfi(cs);
+static struct tu_draw_state
+tu_emit_input_attachments(struct tu_cmd_buffer *cmd,
+ const struct tu_subpass *subpass,
+ bool gmem)
+{
+ /* note: we can probably emit input attachments just once for the whole
+ * renderpass, this would avoid emitting both sysmem/gmem versions
+ *
+ * emit two texture descriptors for each input, as a workaround for
+ * d24s8, which can be sampled as both float (depth) and integer (stencil)
+ * tu_shader lowers uint input attachment loads to use the 2nd descriptor
+ * in the pair
+ * TODO: a smarter workaround
+ */
- tu_cs_emit_regs(cs,
- A6XX_RB_CCU_CNTL(.unknown = phys_dev->magic.RB_CCU_CNTL_gmem));
+ if (!subpass->input_count)
+ return (struct tu_draw_state) {};
- cmd->wait_for_idle = false;
-}
+ struct tu_cs_memory texture;
+ VkResult result = tu_cs_alloc(&cmd->sub_cs, subpass->input_count * 2,
+ A6XX_TEX_CONST_DWORDS, &texture);
+ assert(result == VK_SUCCESS);
-static void
-tu_emit_sysmem_clear_attachment(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
- uint32_t a,
- const VkRenderPassBeginInfo *info)
-{
- const struct tu_framebuffer *fb = cmd->state.framebuffer;
- const struct tu_image_view *iview = fb->attachments[a].attachment;
- const struct tu_render_pass_attachment *attachment =
- &cmd->state.pass->attachments[a];
- unsigned clear_mask = 0;
+ for (unsigned i = 0; i < subpass->input_count * 2; i++) {
+ uint32_t a = subpass->input_attachments[i / 2].attachment;
+ if (a == VK_ATTACHMENT_UNUSED)
+ continue;
- /* note: this means it isn't used by any subpass and shouldn't be cleared anyway */
- if (attachment->gmem_offset < 0)
- return;
+ struct tu_image_view *iview =
+ cmd->state.framebuffer->attachments[a].attachment;
+ const struct tu_render_pass_attachment *att =
+ &cmd->state.pass->attachments[a];
+ uint32_t *dst = &texture.map[A6XX_TEX_CONST_DWORDS * i];
- if (attachment->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
- clear_mask = 0xf;
- }
+ memcpy(dst, iview->descriptor, A6XX_TEX_CONST_DWORDS * 4);
+
+ if (i % 2 == 1 && att->format == VK_FORMAT_D24_UNORM_S8_UINT) {
+ /* note this works because spec says fb and input attachments
+ * must use identity swizzle
+ */
+ dst[0] &= ~(A6XX_TEX_CONST_0_FMT__MASK |
+ A6XX_TEX_CONST_0_SWIZ_X__MASK | A6XX_TEX_CONST_0_SWIZ_Y__MASK |
+ A6XX_TEX_CONST_0_SWIZ_Z__MASK | A6XX_TEX_CONST_0_SWIZ_W__MASK);
+ dst[0] |= A6XX_TEX_CONST_0_FMT(FMT6_S8Z24_UINT) |
+ A6XX_TEX_CONST_0_SWIZ_X(A6XX_TEX_Y) |
+ A6XX_TEX_CONST_0_SWIZ_Y(A6XX_TEX_ZERO) |
+ A6XX_TEX_CONST_0_SWIZ_Z(A6XX_TEX_ZERO) |
+ A6XX_TEX_CONST_0_SWIZ_W(A6XX_TEX_ONE);
+ }
+
+ if (!gmem)
+ continue;
- if (vk_format_has_stencil(iview->vk_format)) {
- clear_mask &= 0x1;
- if (attachment->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR)
- clear_mask |= 0x2;
- if (clear_mask != 0x3)
- tu_finishme("depth/stencil only load op");
+ /* patched for gmem */
+ dst[0] &= ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
+ dst[0] |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
+ dst[2] =
+ A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D) |
+ A6XX_TEX_CONST_2_PITCH(cmd->state.framebuffer->tile0.width * att->cpp);
+ dst[3] = 0;
+ dst[4] = cmd->device->physical_device->gmem_base + att->gmem_offset;
+ dst[5] = A6XX_TEX_CONST_5_DEPTH(1);
+ for (unsigned i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
+ dst[i] = 0;
}
- if (!clear_mask)
- return;
+ struct tu_cs cs;
+ struct tu_draw_state ds = tu_cs_draw_state(&cmd->sub_cs, &cs, 9);
+
+ tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_FRAG, 3);
+ tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(SB6_FS_TEX) |
+ CP_LOAD_STATE6_0_NUM_UNIT(subpass->input_count * 2));
+ tu_cs_emit_qw(&cs, texture.iova);
+
+ tu_cs_emit_pkt4(&cs, REG_A6XX_SP_FS_TEX_CONST_LO, 2);
+ tu_cs_emit_qw(&cs, texture.iova);
+
+ tu_cs_emit_regs(&cs, A6XX_SP_FS_TEX_COUNT(subpass->input_count * 2));
- tu_clear_sysmem_attachment(cmd, cs, a,
- &info->pClearValues[a], &(struct VkClearRect) {
- .rect = info->renderArea,
- .baseArrayLayer = iview->base_layer,
- .layerCount = iview->layer_count,
- });
+ assert(cs.cur == cs.end); /* validate draw state size */
+
+ return ds;
+}
+
+static void
+tu_set_input_attachments(struct tu_cmd_buffer *cmd, const struct tu_subpass *subpass)
+{
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 6);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_INPUT_ATTACHMENTS_GMEM,
+ tu_emit_input_attachments(cmd, subpass, true));
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_INPUT_ATTACHMENTS_SYSMEM,
+ tu_emit_input_attachments(cmd, subpass, false));
}
static void
-tu_emit_load_clear(struct tu_cmd_buffer *cmd,
- const VkRenderPassBeginInfo *info)
+tu_emit_renderpass_begin(struct tu_cmd_buffer *cmd,
+ const VkRenderPassBeginInfo *info)
{
struct tu_cs *cs = &cmd->draw_cs;
tu6_emit_blit_scissor(cmd, cs, true);
for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
- tu6_emit_load_attachment(cmd, cs, i);
+ tu_load_gmem_attachment(cmd, cs, i, false);
tu6_emit_blit_scissor(cmd, cs, false);
for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
- tu6_emit_clear_attachment(cmd, cs, i, info);
+ tu_clear_gmem_attachment(cmd, cs, i, info);
tu_cond_exec_end(cs);
- /* invalidate because reading input attachments will cache GMEM and
- * the cache isn''t updated when GMEM is written
- * TODO: is there a no-cache bit for textures?
- */
- if (cmd->state.subpass->input_count)
- tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
-
tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
- tu_emit_sysmem_clear_attachment(cmd, cs, i, info);
+ tu_clear_sysmem_attachment(cmd, cs, i, info);
tu_cond_exec_end(cs);
}
static void
-tu6_sysmem_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
- const struct VkRect2D *renderArea)
+tu6_sysmem_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
const struct tu_framebuffer *fb = cmd->state.framebuffer;
assert(fb->width > 0 && fb->height > 0);
- tu6_emit_window_scissor(cmd, cs, 0, 0, fb->width - 1, fb->height - 1);
- tu6_emit_window_offset(cmd, cs, 0, 0);
+ tu6_emit_window_scissor(cs, 0, 0, fb->width - 1, fb->height - 1);
+ tu6_emit_window_offset(cs, 0, 0);
tu6_emit_bin_size(cs, 0, 0, 0xc00000); /* 0xc00000 = BYPASS? */
- tu6_emit_lrz_flush(cmd, cs);
+ tu6_emit_event_write(cmd, cs, LRZ_FLUSH);
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_BYPASS));
tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
tu_cs_emit(cs, 0x0);
- tu6_emit_event_write(cmd, cs, PC_CCU_INVALIDATE_COLOR, false);
- tu6_emit_event_write(cmd, cs, PC_CCU_INVALIDATE_DEPTH, false);
- tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
-
- tu6_emit_wfi(cmd, cs);
- tu_cs_emit_regs(cs,
- A6XX_RB_CCU_CNTL(0x10000000));
+ tu_emit_cache_flush_ccu(cmd, cs, TU_CMD_CCU_SYSMEM);
/* enable stream-out, with sysmem there is only one pass: */
- tu_cs_emit_regs(cs,
- A6XX_VPC_SO_OVERRIDE(.so_disable = false));
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(false));
tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
tu_cs_emit(cs, 0x1);
/* Do any resolves of the last subpass. These are handled in the
* tile_store_ib in the gmem path.
*/
-
- const struct tu_subpass *subpass = cmd->state.subpass;
- if (subpass->resolve_attachments) {
- for (unsigned i = 0; i < subpass->color_count; i++) {
- uint32_t a = subpass->resolve_attachments[i].attachment;
- if (a != VK_ATTACHMENT_UNUSED)
- tu6_emit_sysmem_resolve(cmd, cs, a,
- subpass->color_attachments[i].attachment);
- }
- }
+ tu6_emit_sysmem_resolves(cmd, cs, cmd->state.subpass);
tu_cs_emit_call(cs, &cmd->draw_epilogue_cs);
tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
tu_cs_emit(cs, 0x0);
- tu6_emit_lrz_flush(cmd, cs);
-
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS, true);
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_DEPTH_TS, true);
+ tu6_emit_event_write(cmd, cs, LRZ_FLUSH);
tu_cs_sanity_check(cs);
}
-
static void
tu6_tile_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
struct tu_physical_device *phys_dev = cmd->device->physical_device;
- tu6_emit_lrz_flush(cmd, cs);
+ tu6_emit_event_write(cmd, cs, LRZ_FLUSH);
/* lrz clear? */
- tu6_emit_cache_flush(cmd, cs);
-
tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
tu_cs_emit(cs, 0x0);
- /* 0x10000000 for BYPASS.. 0x7c13c080 for GMEM: */
- tu6_emit_wfi(cmd, cs);
- tu_cs_emit_regs(cs,
- A6XX_RB_CCU_CNTL(phys_dev->magic.RB_CCU_CNTL_gmem));
+ tu_emit_cache_flush_ccu(cmd, cs, TU_CMD_CCU_GMEM);
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
if (use_hw_binning(cmd)) {
/* enable stream-out during binning pass: */
- tu_cs_emit_regs(cs, A6XX_VPC_SO_OVERRIDE(.so_disable=false));
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(false));
- tu6_emit_bin_size(cs,
- tiling->tile0.extent.width,
- tiling->tile0.extent.height,
+ tu6_emit_bin_size(cs, fb->tile0.width, fb->tile0.height,
A6XX_RB_BIN_CONTROL_BINNING_PASS | 0x6000000);
tu6_emit_render_cntl(cmd, cmd->state.subpass, cs, true);
tu6_emit_binning_pass(cmd, cs);
/* and disable stream-out for draw pass: */
- tu_cs_emit_regs(cs, A6XX_VPC_SO_OVERRIDE(.so_disable=true));
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(true));
- tu6_emit_bin_size(cs,
- tiling->tile0.extent.width,
- tiling->tile0.extent.height,
+ tu6_emit_bin_size(cs, fb->tile0.width, fb->tile0.height,
A6XX_RB_BIN_CONTROL_USE_VIZ | 0x6000000);
tu_cs_emit_regs(cs,
tu_cs_emit(cs, 0x1);
} else {
/* no binning pass, so enable stream-out for draw pass:: */
- tu_cs_emit_regs(cs, A6XX_VPC_SO_OVERRIDE(.so_disable=false));
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_DISABLE(false));
- tu6_emit_bin_size(cs,
- tiling->tile0.extent.width,
- tiling->tile0.extent.height,
- 0x6000000);
+ tu6_emit_bin_size(cs, fb->tile0.width, fb->tile0.height, 0x6000000);
}
tu_cs_sanity_check(cs);
}
static void
-tu6_render_tile(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_tile *tile)
+tu6_render_tile(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
- tu6_emit_tile_select(cmd, cs, tile);
-
tu_cs_emit_call(cs, &cmd->draw_cs);
- cmd->wait_for_idle = true;
if (use_hw_binning(cmd)) {
- tu_cs_emit_pkt7(cs, CP_REG_TEST, 1);
- tu_cs_emit(cs, A6XX_CP_REG_TEST_0_REG(OVERFLOW_FLAG_REG) |
- A6XX_CP_REG_TEST_0_BIT(0) |
- A6XX_CP_REG_TEST_0_WAIT_FOR_ME);
-
- tu_cs_reserve(cs, 3 + 2);
- tu_cs_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
- tu_cs_emit(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
- tu_cs_emit(cs, CP_COND_REG_EXEC_1_DWORDS(2));
-
- /* if (no overflow) */ {
- tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
- tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_ENDVIS));
- }
+ tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
+ tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_ENDVIS));
}
tu_cs_emit_ib(cs, &cmd->state.tile_store_ib);
tu_cs_emit_regs(cs,
A6XX_GRAS_LRZ_CNTL(0));
- tu6_emit_lrz_flush(cmd, cs);
+ tu6_emit_event_write(cmd, cs, LRZ_FLUSH);
- tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS, true);
+ tu6_emit_event_write(cmd, cs, PC_CCU_RESOLVE_TS);
tu_cs_sanity_check(cs);
}
static void
tu_cmd_render_tiles(struct tu_cmd_buffer *cmd)
{
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
tu6_tile_render_begin(cmd, &cmd->cs);
- for (uint32_t y = 0; y < tiling->tile_count.height; y++) {
- for (uint32_t x = 0; x < tiling->tile_count.width; x++) {
- struct tu_tile tile;
- tu_tiling_config_get_tile(tiling, cmd->device, x, y, &tile);
- tu6_render_tile(cmd, &cmd->cs, &tile);
+ uint32_t pipe = 0;
+ for (uint32_t py = 0; py < fb->pipe_count.height; py++) {
+ for (uint32_t px = 0; px < fb->pipe_count.width; px++, pipe++) {
+ uint32_t tx1 = px * fb->pipe0.width;
+ uint32_t ty1 = py * fb->pipe0.height;
+ uint32_t tx2 = MIN2(tx1 + fb->pipe0.width, fb->tile_count.width);
+ uint32_t ty2 = MIN2(ty1 + fb->pipe0.height, fb->tile_count.height);
+ uint32_t slot = 0;
+ for (uint32_t ty = ty1; ty < ty2; ty++) {
+ for (uint32_t tx = tx1; tx < tx2; tx++, slot++) {
+ tu6_emit_tile_select(cmd, &cmd->cs, tx, ty, pipe, slot);
+ tu6_render_tile(cmd, &cmd->cs);
+ }
+ }
}
}
static void
tu_cmd_render_sysmem(struct tu_cmd_buffer *cmd)
{
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
-
- tu6_sysmem_render_begin(cmd, &cmd->cs, &tiling->render_area);
+ tu6_sysmem_render_begin(cmd, &cmd->cs);
tu_cs_emit_call(&cmd->cs, &cmd->draw_cs);
- cmd->wait_for_idle = true;
tu6_sysmem_render_end(cmd, &cmd->cs);
}
static void
tu_cmd_prepare_tile_store_ib(struct tu_cmd_buffer *cmd)
{
- const uint32_t tile_store_space = 32 + 23 * cmd->state.pass->attachment_count;
+ const uint32_t tile_store_space = 11 + (35 * 2) * cmd->state.pass->attachment_count;
struct tu_cs sub_cs;
VkResult result =
cmd->state.tile_store_ib = tu_cs_end_sub_stream(&cmd->sub_cs, &sub_cs);
}
-static void
-tu_cmd_update_tiling_config(struct tu_cmd_buffer *cmd,
- const VkRect2D *render_area)
-{
- const struct tu_device *dev = cmd->device;
- struct tu_tiling_config *tiling = &cmd->state.tiling_config;
-
- tiling->render_area = *render_area;
- tiling->force_sysmem = force_sysmem(cmd, render_area);
-
- tu_tiling_config_update_tile_layout(tiling, dev, cmd->state.pass->gmem_pixels);
- tu_tiling_config_update_pipe_layout(tiling, dev);
- tu_tiling_config_update_pipes(tiling, dev);
-}
-
-const struct tu_dynamic_state default_dynamic_state = {
- .viewport =
- {
- .count = 0,
- },
- .scissor =
- {
- .count = 0,
- },
- .line_width = 1.0f,
- .depth_bias =
- {
- .bias = 0.0f,
- .clamp = 0.0f,
- .slope = 0.0f,
- },
- .blend_constants = { 0.0f, 0.0f, 0.0f, 0.0f },
- .depth_bounds =
- {
- .min = 0.0f,
- .max = 1.0f,
- },
- .stencil_compare_mask =
- {
- .front = ~0u,
- .back = ~0u,
- },
- .stencil_write_mask =
- {
- .front = ~0u,
- .back = ~0u,
- },
- .stencil_reference =
- {
- .front = 0u,
- .back = 0u,
- },
-};
-
-static void UNUSED /* FINISHME */
-tu_bind_dynamic_state(struct tu_cmd_buffer *cmd_buffer,
- const struct tu_dynamic_state *src)
-{
- struct tu_dynamic_state *dest = &cmd_buffer->state.dynamic;
- uint32_t copy_mask = src->mask;
- uint32_t dest_mask = 0;
-
- tu_use_args(cmd_buffer); /* FINISHME */
-
- /* Make sure to copy the number of viewports/scissors because they can
- * only be specified at pipeline creation time.
- */
- dest->viewport.count = src->viewport.count;
- dest->scissor.count = src->scissor.count;
- dest->discard_rectangle.count = src->discard_rectangle.count;
-
- if (copy_mask & TU_DYNAMIC_VIEWPORT) {
- if (memcmp(&dest->viewport.viewports, &src->viewport.viewports,
- src->viewport.count * sizeof(VkViewport))) {
- typed_memcpy(dest->viewport.viewports, src->viewport.viewports,
- src->viewport.count);
- dest_mask |= TU_DYNAMIC_VIEWPORT;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_SCISSOR) {
- if (memcmp(&dest->scissor.scissors, &src->scissor.scissors,
- src->scissor.count * sizeof(VkRect2D))) {
- typed_memcpy(dest->scissor.scissors, src->scissor.scissors,
- src->scissor.count);
- dest_mask |= TU_DYNAMIC_SCISSOR;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_LINE_WIDTH) {
- if (dest->line_width != src->line_width) {
- dest->line_width = src->line_width;
- dest_mask |= TU_DYNAMIC_LINE_WIDTH;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_DEPTH_BIAS) {
- if (memcmp(&dest->depth_bias, &src->depth_bias,
- sizeof(src->depth_bias))) {
- dest->depth_bias = src->depth_bias;
- dest_mask |= TU_DYNAMIC_DEPTH_BIAS;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_BLEND_CONSTANTS) {
- if (memcmp(&dest->blend_constants, &src->blend_constants,
- sizeof(src->blend_constants))) {
- typed_memcpy(dest->blend_constants, src->blend_constants, 4);
- dest_mask |= TU_DYNAMIC_BLEND_CONSTANTS;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_DEPTH_BOUNDS) {
- if (memcmp(&dest->depth_bounds, &src->depth_bounds,
- sizeof(src->depth_bounds))) {
- dest->depth_bounds = src->depth_bounds;
- dest_mask |= TU_DYNAMIC_DEPTH_BOUNDS;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_STENCIL_COMPARE_MASK) {
- if (memcmp(&dest->stencil_compare_mask, &src->stencil_compare_mask,
- sizeof(src->stencil_compare_mask))) {
- dest->stencil_compare_mask = src->stencil_compare_mask;
- dest_mask |= TU_DYNAMIC_STENCIL_COMPARE_MASK;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_STENCIL_WRITE_MASK) {
- if (memcmp(&dest->stencil_write_mask, &src->stencil_write_mask,
- sizeof(src->stencil_write_mask))) {
- dest->stencil_write_mask = src->stencil_write_mask;
- dest_mask |= TU_DYNAMIC_STENCIL_WRITE_MASK;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_STENCIL_REFERENCE) {
- if (memcmp(&dest->stencil_reference, &src->stencil_reference,
- sizeof(src->stencil_reference))) {
- dest->stencil_reference = src->stencil_reference;
- dest_mask |= TU_DYNAMIC_STENCIL_REFERENCE;
- }
- }
-
- if (copy_mask & TU_DYNAMIC_DISCARD_RECTANGLE) {
- if (memcmp(&dest->discard_rectangle.rectangles,
- &src->discard_rectangle.rectangles,
- src->discard_rectangle.count * sizeof(VkRect2D))) {
- typed_memcpy(dest->discard_rectangle.rectangles,
- src->discard_rectangle.rectangles,
- src->discard_rectangle.count);
- dest_mask |= TU_DYNAMIC_DISCARD_RECTANGLE;
- }
- }
-}
-
static VkResult
tu_create_cmd_buffer(struct tu_device *device,
struct tu_cmd_pool *pool,
VkCommandBuffer *pCommandBuffer)
{
struct tu_cmd_buffer *cmd_buffer;
- cmd_buffer = vk_zalloc(&pool->alloc, sizeof(*cmd_buffer), 8,
- VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+
+ cmd_buffer = vk_object_zalloc(&device->vk, NULL, sizeof(*cmd_buffer),
+ VK_OBJECT_TYPE_COMMAND_BUFFER);
if (cmd_buffer == NULL)
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
- cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
cmd_buffer->device = device;
cmd_buffer->pool = pool;
cmd_buffer->level = level;
list_inithead(&cmd_buffer->upload.list);
- VkResult result = tu_bo_init_new(device, &cmd_buffer->scratch_bo, 0x1000);
- if (result != VK_SUCCESS)
- goto fail_scratch_bo;
-
- /* TODO: resize on overflow */
- cmd_buffer->vsc_data_pitch = device->vsc_data_pitch;
- cmd_buffer->vsc_data2_pitch = device->vsc_data2_pitch;
- cmd_buffer->vsc_data = device->vsc_data;
- cmd_buffer->vsc_data2 = device->vsc_data2;
-
return VK_SUCCESS;
-
-fail_scratch_bo:
- list_del(&cmd_buffer->pool_link);
- return result;
}
static void
tu_cmd_buffer_destroy(struct tu_cmd_buffer *cmd_buffer)
{
- tu_bo_finish(cmd_buffer->device, &cmd_buffer->scratch_bo);
-
list_del(&cmd_buffer->pool_link);
- for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++)
- free(cmd_buffer->descriptors[i].push_set.set.mapped_ptr);
-
tu_cs_finish(&cmd_buffer->cs);
tu_cs_finish(&cmd_buffer->draw_cs);
tu_cs_finish(&cmd_buffer->draw_epilogue_cs);
tu_cs_finish(&cmd_buffer->sub_cs);
tu_bo_list_destroy(&cmd_buffer->bo_list);
- vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
+ vk_object_free(&cmd_buffer->device->vk, &cmd_buffer->pool->alloc, cmd_buffer);
}
static VkResult
tu_reset_cmd_buffer(struct tu_cmd_buffer *cmd_buffer)
{
- cmd_buffer->wait_for_idle = true;
-
cmd_buffer->record_result = VK_SUCCESS;
tu_bo_list_reset(&cmd_buffer->bo_list);
tu_cs_reset(&cmd_buffer->draw_epilogue_cs);
tu_cs_reset(&cmd_buffer->sub_cs);
- for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++) {
- cmd_buffer->descriptors[i].valid = 0;
- cmd_buffer->descriptors[i].push_dirty = false;
- }
+ for (unsigned i = 0; i < MAX_BIND_POINTS; i++)
+ memset(&cmd_buffer->descriptors[i].sets, 0, sizeof(cmd_buffer->descriptors[i].sets));
cmd_buffer->status = TU_CMD_BUFFER_STATUS_INITIAL;
list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
result = tu_reset_cmd_buffer(cmd_buffer);
- cmd_buffer->_loader_data.loaderMagic = ICD_LOADER_MAGIC;
cmd_buffer->level = pAllocateInfo->level;
pCommandBuffers[i] = tu_cmd_buffer_to_handle(cmd_buffer);
return tu_reset_cmd_buffer(cmd_buffer);
}
+/* Initialize the cache, assuming all necessary flushes have happened but *not*
+ * invalidations.
+ */
+static void
+tu_cache_init(struct tu_cache_state *cache)
+{
+ cache->flush_bits = 0;
+ cache->pending_flush_bits = TU_CMD_FLAG_ALL_INVALIDATE;
+}
+
VkResult
tu_BeginCommandBuffer(VkCommandBuffer commandBuffer,
const VkCommandBufferBeginInfo *pBeginInfo)
}
memset(&cmd_buffer->state, 0, sizeof(cmd_buffer->state));
+ cmd_buffer->state.index_size = 0xff; /* dirty restart index */
+
+ tu_cache_init(&cmd_buffer->state.cache);
+ tu_cache_init(&cmd_buffer->state.renderpass_cache);
cmd_buffer->usage_flags = pBeginInfo->flags;
tu_cs_begin(&cmd_buffer->cs);
tu_cs_begin(&cmd_buffer->draw_cs);
tu_cs_begin(&cmd_buffer->draw_epilogue_cs);
- cmd_buffer->scratch_seqno = 0;
-
/* setup initial configuration into command buffer */
if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
switch (cmd_buffer->queue_family_index) {
default:
break;
}
- } else if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY &&
- (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT)) {
- assert(pBeginInfo->pInheritanceInfo);
- cmd_buffer->state.pass = tu_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
- cmd_buffer->state.subpass = &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
+ } else if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY) {
+ if (pBeginInfo->flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
+ assert(pBeginInfo->pInheritanceInfo);
+ cmd_buffer->state.pass = tu_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
+ cmd_buffer->state.subpass =
+ &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];
+ } else {
+ /* When executing in the middle of another command buffer, the CCU
+ * state is unknown.
+ */
+ cmd_buffer->state.ccu_state = TU_CMD_CCU_UNKNOWN;
+ }
}
cmd_buffer->status = TU_CMD_BUFFER_STATUS_RECORDING;
return VK_SUCCESS;
}
+/* Sets vertex buffers to HW binding points. We emit VBs in SDS (so that bin
+ * rendering can skip over unused state), so we need to collect all the
+ * bindings together into a single state emit at draw time.
+ */
void
tu_CmdBindVertexBuffers(VkCommandBuffer commandBuffer,
uint32_t firstBinding,
assert(firstBinding + bindingCount <= MAX_VBS);
for (uint32_t i = 0; i < bindingCount; i++) {
- cmd->state.vb.buffers[firstBinding + i] =
- tu_buffer_from_handle(pBuffers[i]);
+ struct tu_buffer *buf = tu_buffer_from_handle(pBuffers[i]);
+
+ cmd->state.vb.buffers[firstBinding + i] = buf;
cmd->state.vb.offsets[firstBinding + i] = pOffsets[i];
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
}
- /* VB states depend on VkPipelineVertexInputStateCreateInfo */
cmd->state.dirty |= TU_CMD_DIRTY_VERTEX_BUFFERS;
}
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_buffer, buf, buffer);
- /* initialize/update the restart index */
- if (!cmd->state.index_buffer || cmd->state.index_type != indexType) {
- struct tu_cs *draw_cs = &cmd->draw_cs;
- tu6_emit_restart_index(
- draw_cs, indexType == VK_INDEX_TYPE_UINT32 ? 0xffffffff : 0xffff);
- tu_cs_sanity_check(draw_cs);
+ uint32_t index_size, index_shift, restart_index;
+
+ switch (indexType) {
+ case VK_INDEX_TYPE_UINT16:
+ index_size = INDEX4_SIZE_16_BIT;
+ index_shift = 1;
+ restart_index = 0xffff;
+ break;
+ case VK_INDEX_TYPE_UINT32:
+ index_size = INDEX4_SIZE_32_BIT;
+ index_shift = 2;
+ restart_index = 0xffffffff;
+ break;
+ case VK_INDEX_TYPE_UINT8_EXT:
+ index_size = INDEX4_SIZE_8_BIT;
+ index_shift = 0;
+ restart_index = 0xff;
+ break;
+ default:
+ unreachable("invalid VkIndexType");
}
- /* track the BO */
- if (cmd->state.index_buffer != buf)
- tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
+ /* initialize/update the restart index */
+ if (cmd->state.index_size != index_size)
+ tu_cs_emit_regs(&cmd->draw_cs, A6XX_PC_RESTART_INDEX(restart_index));
+
+ assert(buf->size >= offset);
- cmd->state.index_buffer = buf;
- cmd->state.index_offset = offset;
- cmd->state.index_type = indexType;
+ cmd->state.index_va = buf->bo->iova + buf->bo_offset + offset;
+ cmd->state.max_index_count = (buf->size - offset) >> index_shift;
+ cmd->state.index_size = index_size;
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
}
void
uint32_t dynamicOffsetCount,
const uint32_t *pDynamicOffsets)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_pipeline_layout, layout, _layout);
unsigned dyn_idx = 0;
struct tu_descriptor_state *descriptors_state =
- tu_get_descriptors_state(cmd_buffer, pipelineBindPoint);
+ tu_get_descriptors_state(cmd, pipelineBindPoint);
for (unsigned i = 0; i < descriptorSetCount; ++i) {
unsigned idx = i + firstSet;
TU_FROM_HANDLE(tu_descriptor_set, set, pDescriptorSets[i]);
descriptors_state->sets[idx] = set;
- descriptors_state->valid |= (1u << idx);
for(unsigned j = 0; j < set->layout->dynamic_offset_count; ++j, ++dyn_idx) {
- unsigned idx = j + layout->set[i + firstSet].dynamic_offset_start;
+ /* update the contents of the dynamic descriptor set */
+ unsigned src_idx = j;
+ unsigned dst_idx = j + layout->set[idx].dynamic_offset_start;
assert(dyn_idx < dynamicOffsetCount);
- descriptors_state->dynamic_buffers[idx] =
- set->dynamic_descriptors[j].va + pDynamicOffsets[dyn_idx];
+ uint32_t *dst =
+ &descriptors_state->dynamic_descriptors[dst_idx * A6XX_TEX_CONST_DWORDS];
+ uint32_t *src =
+ &set->dynamic_descriptors[src_idx * A6XX_TEX_CONST_DWORDS];
+ uint32_t offset = pDynamicOffsets[dyn_idx];
+
+ /* Patch the storage/uniform descriptors right away. */
+ if (layout->set[idx].layout->dynamic_ubo & (1 << j)) {
+ /* Note: we can assume here that the addition won't roll over and
+ * change the SIZE field.
+ */
+ uint64_t va = src[0] | ((uint64_t)src[1] << 32);
+ va += offset;
+ dst[0] = va;
+ dst[1] = va >> 32;
+ } else {
+ memcpy(dst, src, A6XX_TEX_CONST_DWORDS * 4);
+ /* Note: A6XX_IBO_5_DEPTH is always 0 */
+ uint64_t va = dst[4] | ((uint64_t)dst[5] << 32);
+ va += offset;
+ dst[4] = va;
+ dst[5] = va >> 32;
+ }
+ }
+
+ for (unsigned j = 0; j < set->layout->buffer_count; ++j) {
+ if (set->buffers[j]) {
+ tu_bo_list_add(&cmd->bo_list, set->buffers[j],
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
+ }
+ }
+
+ if (set->size > 0) {
+ tu_bo_list_add(&cmd->bo_list, &set->pool->bo,
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
}
}
+ assert(dyn_idx == dynamicOffsetCount);
+
+ uint32_t sp_bindless_base_reg, hlsq_bindless_base_reg, hlsq_invalidate_value;
+ uint64_t addr[MAX_SETS + 1] = {};
+ struct tu_cs *cs, state_cs;
+
+ for (uint32_t i = 0; i < MAX_SETS; i++) {
+ struct tu_descriptor_set *set = descriptors_state->sets[i];
+ if (set)
+ addr[i] = set->va | 3;
+ }
+
+ if (layout->dynamic_offset_count) {
+ /* allocate and fill out dynamic descriptor set */
+ struct tu_cs_memory dynamic_desc_set;
+ VkResult result = tu_cs_alloc(&cmd->sub_cs, layout->dynamic_offset_count,
+ A6XX_TEX_CONST_DWORDS, &dynamic_desc_set);
+ assert(result == VK_SUCCESS);
+
+ memcpy(dynamic_desc_set.map, descriptors_state->dynamic_descriptors,
+ layout->dynamic_offset_count * A6XX_TEX_CONST_DWORDS * 4);
+ addr[MAX_SETS] = dynamic_desc_set.iova | 3;
+ }
+
+ if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS) {
+ sp_bindless_base_reg = REG_A6XX_SP_BINDLESS_BASE(0);
+ hlsq_bindless_base_reg = REG_A6XX_HLSQ_BINDLESS_BASE(0);
+ hlsq_invalidate_value = A6XX_HLSQ_INVALIDATE_CMD_GFX_BINDLESS(0x1f);
+
+ cmd->state.desc_sets = tu_cs_draw_state(&cmd->sub_cs, &state_cs, 24);
+ cmd->state.dirty |= TU_CMD_DIRTY_DESC_SETS_LOAD | TU_CMD_DIRTY_SHADER_CONSTS;
+ cs = &state_cs;
+ } else {
+ assert(pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE);
+
+ sp_bindless_base_reg = REG_A6XX_SP_CS_BINDLESS_BASE(0);
+ hlsq_bindless_base_reg = REG_A6XX_HLSQ_CS_BINDLESS_BASE(0);
+ hlsq_invalidate_value = A6XX_HLSQ_INVALIDATE_CMD_CS_BINDLESS(0x1f);
+
+ cmd->state.dirty |= TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD;
+ cs = &cmd->cs;
+ }
+
+ tu_cs_emit_pkt4(cs, sp_bindless_base_reg, 10);
+ tu_cs_emit_array(cs, (const uint32_t*) addr, 10);
+ tu_cs_emit_pkt4(cs, hlsq_bindless_base_reg, 10);
+ tu_cs_emit_array(cs, (const uint32_t*) addr, 10);
+ tu_cs_emit_regs(cs, A6XX_HLSQ_INVALIDATE_CMD(.dword = hlsq_invalidate_value));
- cmd_buffer->state.dirty |= TU_CMD_DIRTY_DESCRIPTOR_SETS;
+ if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS) {
+ assert(cs->cur == cs->end); /* validate draw state size */
+ tu_cs_emit_pkt7(&cmd->draw_cs, CP_SET_DRAW_STATE, 3);
+ tu_cs_emit_draw_state(&cmd->draw_cs, TU_DRAW_STATE_DESC_SETS, cmd->state.desc_sets);
+ }
}
void tu_CmdBindTransformFeedbackBuffersEXT(VkCommandBuffer commandBuffer,
const VkDeviceSize *pSizes)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- assert(firstBinding + bindingCount <= IR3_MAX_SO_BUFFERS);
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ /* using COND_REG_EXEC for xfb commands matches the blob behavior
+ * presumably there isn't any benefit using a draw state when the
+ * condition is (SYSMEM | BINNING)
+ */
+ tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+ CP_COND_REG_EXEC_0_SYSMEM |
+ CP_COND_REG_EXEC_0_BINNING);
for (uint32_t i = 0; i < bindingCount; i++) {
- uint32_t idx = firstBinding + i;
TU_FROM_HANDLE(tu_buffer, buf, pBuffers[i]);
+ uint64_t iova = buf->bo->iova + pOffsets[i];
+ uint32_t size = buf->bo->size - pOffsets[i];
+ uint32_t idx = i + firstBinding;
- if (pOffsets[i] != 0)
- cmd->state.streamout_reset |= 1 << idx;
+ if (pSizes && pSizes[i] != VK_WHOLE_SIZE)
+ size = pSizes[i];
- cmd->state.streamout_buf.buffers[idx] = buf;
- cmd->state.streamout_buf.offsets[idx] = pOffsets[i];
- cmd->state.streamout_buf.sizes[idx] = pSizes[i];
+ /* BUFFER_BASE is 32-byte aligned, add remaining offset to BUFFER_OFFSET */
+ uint32_t offset = iova & 0x1f;
+ iova &= ~(uint64_t) 0x1f;
- cmd->state.streamout_enabled |= 1 << idx;
+ tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_BUFFER_BASE(idx), 3);
+ tu_cs_emit_qw(cs, iova);
+ tu_cs_emit(cs, size + offset);
+
+ cmd->state.streamout_offset[idx] = offset;
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_WRITE);
}
- cmd->state.dirty |= TU_CMD_DIRTY_STREAMOUT_BUFFERS;
+ tu_cond_exec_end(cs);
}
-void tu_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer,
- uint32_t firstCounterBuffer,
- uint32_t counterBufferCount,
- const VkBuffer *pCounterBuffers,
- const VkDeviceSize *pCounterBufferOffsets)
+void
+tu_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer,
+ uint32_t firstCounterBuffer,
+ uint32_t counterBufferCount,
+ const VkBuffer *pCounterBuffers,
+ const VkDeviceSize *pCounterBufferOffsets)
{
- assert(firstCounterBuffer + counterBufferCount <= IR3_MAX_SO_BUFFERS);
- /* TODO do something with counter buffer? */
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+ CP_COND_REG_EXEC_0_SYSMEM |
+ CP_COND_REG_EXEC_0_BINNING);
+
+ /* TODO: only update offset for active buffers */
+ for (uint32_t i = 0; i < IR3_MAX_SO_BUFFERS; i++)
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_OFFSET(i, cmd->state.streamout_offset[i]));
+
+ for (uint32_t i = 0; i < counterBufferCount; i++) {
+ uint32_t idx = firstCounterBuffer + i;
+ uint32_t offset = cmd->state.streamout_offset[idx];
+
+ if (!pCounterBuffers[i])
+ continue;
+
+ TU_FROM_HANDLE(tu_buffer, buf, pCounterBuffers[i]);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
+
+ tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
+ tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(idx)) |
+ CP_MEM_TO_REG_0_UNK31 |
+ CP_MEM_TO_REG_0_CNT(1));
+ tu_cs_emit_qw(cs, buf->bo->iova + pCounterBufferOffsets[i]);
+
+ if (offset) {
+ tu_cs_emit_pkt7(cs, CP_REG_RMW, 3);
+ tu_cs_emit(cs, CP_REG_RMW_0_DST_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(idx)) |
+ CP_REG_RMW_0_SRC1_ADD);
+ tu_cs_emit_qw(cs, 0xffffffff);
+ tu_cs_emit_qw(cs, offset);
+ }
+ }
+
+ tu_cond_exec_end(cs);
}
void tu_CmdEndTransformFeedbackEXT(VkCommandBuffer commandBuffer,
const VkBuffer *pCounterBuffers,
const VkDeviceSize *pCounterBufferOffsets)
{
- assert(firstCounterBuffer + counterBufferCount <= IR3_MAX_SO_BUFFERS);
- /* TODO do something with counter buffer? */
-
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- cmd->state.streamout_enabled = 0;
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+ CP_COND_REG_EXEC_0_SYSMEM |
+ CP_COND_REG_EXEC_0_BINNING);
+
+ /* TODO: only flush buffers that need to be flushed */
+ for (uint32_t i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
+ /* note: FLUSH_BASE is always the same, so it could go in init_hw()? */
+ tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_FLUSH_BASE(i), 2);
+ tu_cs_emit_qw(cs, global_iova(cmd, flush_base[i]));
+ tu6_emit_event_write(cmd, cs, FLUSH_SO_0 + i);
+ }
+
+ for (uint32_t i = 0; i < counterBufferCount; i++) {
+ uint32_t idx = firstCounterBuffer + i;
+ uint32_t offset = cmd->state.streamout_offset[idx];
+
+ if (!pCounterBuffers[i])
+ continue;
+
+ TU_FROM_HANDLE(tu_buffer, buf, pCounterBuffers[i]);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_WRITE);
+
+ /* VPC_SO_FLUSH_BASE has dwords counter, but counter should be in bytes */
+ tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
+ tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+ CP_MEM_TO_REG_0_SHIFT_BY_2 |
+ 0x40000 | /* ??? */
+ CP_MEM_TO_REG_0_UNK31 |
+ CP_MEM_TO_REG_0_CNT(1));
+ tu_cs_emit_qw(cs, global_iova(cmd, flush_base[idx]));
+
+ if (offset) {
+ tu_cs_emit_pkt7(cs, CP_REG_RMW, 3);
+ tu_cs_emit(cs, CP_REG_RMW_0_DST_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+ CP_REG_RMW_0_SRC1_ADD);
+ tu_cs_emit_qw(cs, 0xffffffff);
+ tu_cs_emit_qw(cs, -offset);
+ }
+
+ tu_cs_emit_pkt7(cs, CP_REG_TO_MEM, 3);
+ tu_cs_emit(cs, CP_REG_TO_MEM_0_REG(REG_A6XX_CP_SCRATCH_REG(0)) |
+ CP_REG_TO_MEM_0_CNT(1));
+ tu_cs_emit_qw(cs, buf->bo->iova + pCounterBufferOffsets[i]);
+ }
+
+ tu_cond_exec_end(cs);
+
+ cmd->state.xfb_used = true;
}
void
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
memcpy((void*) cmd->push_constants + offset, pValues, size);
- cmd->state.dirty |= TU_CMD_DIRTY_PUSH_CONSTANTS;
+ cmd->state.dirty |= TU_CMD_DIRTY_SHADER_CONSTS;
}
-VkResult
-tu_EndCommandBuffer(VkCommandBuffer commandBuffer)
+/* Flush everything which has been made available but we haven't actually
+ * flushed yet.
+ */
+static void
+tu_flush_all_pending(struct tu_cache_state *cache)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+ cache->flush_bits |= cache->pending_flush_bits & TU_CMD_FLAG_ALL_FLUSH;
+ cache->pending_flush_bits &= ~TU_CMD_FLAG_ALL_FLUSH;
+}
- if (cmd_buffer->scratch_seqno) {
- tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->scratch_bo,
- MSM_SUBMIT_BO_WRITE);
- }
+VkResult
+tu_EndCommandBuffer(VkCommandBuffer commandBuffer)
+{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- if (cmd_buffer->use_vsc_data) {
- tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->vsc_data,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->vsc_data2,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
+ /* We currently flush CCU at the end of the command buffer, like
+ * what the blob does. There's implicit synchronization around every
+ * vkQueueSubmit, but the kernel only flushes the UCHE, and we don't
+ * know yet if this command buffer will be the last in the submit so we
+ * have to defensively flush everything else.
+ *
+ * TODO: We could definitely do better than this, since these flushes
+ * aren't required by Vulkan, but we'd need kernel support to do that.
+ * Ideally, we'd like the kernel to flush everything afterwards, so that we
+ * wouldn't have to do any flushes here, and when submitting multiple
+ * command buffers there wouldn't be any unnecessary flushes in between.
+ */
+ if (cmd_buffer->state.pass) {
+ tu_flush_all_pending(&cmd_buffer->state.renderpass_cache);
+ tu_emit_cache_flush_renderpass(cmd_buffer, &cmd_buffer->draw_cs);
+ } else {
+ tu_flush_all_pending(&cmd_buffer->state.cache);
+ cmd_buffer->state.cache.flush_bits |=
+ TU_CMD_FLAG_CCU_FLUSH_COLOR |
+ TU_CMD_FLAG_CCU_FLUSH_DEPTH;
+ tu_emit_cache_flush(cmd_buffer, &cmd_buffer->cs);
}
- tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->device->border_color,
- MSM_SUBMIT_BO_READ);
+ tu_bo_list_add(&cmd_buffer->bo_list, &cmd_buffer->device->global_bo,
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
for (uint32_t i = 0; i < cmd_buffer->draw_cs.bo_count; i++) {
tu_bo_list_add(&cmd_buffer->bo_list, cmd_buffer->draw_cs.bos[i],
return cmd_buffer->record_result;
}
+static struct tu_cs
+tu_cmd_dynamic_state(struct tu_cmd_buffer *cmd, uint32_t id, uint32_t size)
+{
+ struct tu_cs cs;
+
+ assert(id < ARRAY_SIZE(cmd->state.dynamic_state));
+ cmd->state.dynamic_state[id] = tu_cs_draw_state(&cmd->sub_cs, &cs, size);
+
+ tu_cs_emit_pkt7(&cmd->draw_cs, CP_SET_DRAW_STATE, 3);
+ tu_cs_emit_draw_state(&cmd->draw_cs, TU_DRAW_STATE_DYNAMIC + id, cmd->state.dynamic_state[id]);
+
+ return cs;
+}
+
void
tu_CmdBindPipeline(VkCommandBuffer commandBuffer,
VkPipelineBindPoint pipelineBindPoint,
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_pipeline, pipeline, _pipeline);
- switch (pipelineBindPoint) {
- case VK_PIPELINE_BIND_POINT_GRAPHICS:
- cmd->state.pipeline = pipeline;
- cmd->state.dirty |= TU_CMD_DIRTY_PIPELINE;
- break;
- case VK_PIPELINE_BIND_POINT_COMPUTE:
- cmd->state.compute_pipeline = pipeline;
- cmd->state.dirty |= TU_CMD_DIRTY_COMPUTE_PIPELINE;
- break;
- default:
- unreachable("unrecognized pipeline bind point");
- break;
- }
-
- tu_bo_list_add(&cmd->bo_list, &pipeline->program.binary_bo,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
for (uint32_t i = 0; i < pipeline->cs.bo_count; i++) {
tu_bo_list_add(&cmd->bo_list, pipeline->cs.bos[i],
MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
}
+
+ if (pipelineBindPoint == VK_PIPELINE_BIND_POINT_COMPUTE) {
+ cmd->state.compute_pipeline = pipeline;
+ tu_cs_emit_state_ib(&cmd->cs, pipeline->program.state);
+ return;
+ }
+
+ assert(pipelineBindPoint == VK_PIPELINE_BIND_POINT_GRAPHICS);
+
+ cmd->state.pipeline = pipeline;
+ cmd->state.dirty |= TU_CMD_DIRTY_DESC_SETS_LOAD | TU_CMD_DIRTY_SHADER_CONSTS;
+
+ struct tu_cs *cs = &cmd->draw_cs;
+ uint32_t mask = ~pipeline->dynamic_state_mask & BITFIELD_MASK(TU_DYNAMIC_STATE_COUNT);
+ uint32_t i;
+
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * (7 + util_bitcount(mask)));
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM, pipeline->program.state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM_BINNING, pipeline->program.binning_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VI, pipeline->vi.state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VI_BINNING, pipeline->vi.binning_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_RAST, pipeline->rast_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DS, pipeline->ds_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_BLEND, pipeline->blend_state);
+ for_each_bit(i, mask)
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + i, pipeline->dynamic_state[i]);
+
+ /* If the new pipeline requires more VBs than we had previously set up, we
+ * need to re-emit them in SDS. If it requires the same set or fewer, we
+ * can just re-use the old SDS.
+ */
+ if (pipeline->vi.bindings_used & ~cmd->vertex_bindings_set)
+ cmd->state.dirty |= TU_CMD_DIRTY_VERTEX_BUFFERS;
+
+ /* dynamic linewidth state depends pipeline state's gras_su_cntl
+ * so the dynamic state ib must be updated when pipeline changes
+ */
+ if (pipeline->dynamic_state_mask & BIT(VK_DYNAMIC_STATE_LINE_WIDTH)) {
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_LINE_WIDTH, 2);
+
+ cmd->state.dynamic_gras_su_cntl &= A6XX_GRAS_SU_CNTL_LINEHALFWIDTH__MASK;
+ cmd->state.dynamic_gras_su_cntl |= pipeline->gras_su_cntl;
+
+ tu_cs_emit_regs(&cs, A6XX_GRAS_SU_CNTL(.dword = cmd->state.dynamic_gras_su_cntl));
+ }
}
void
const VkViewport *pViewports)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- struct tu_cs *draw_cs = &cmd->draw_cs;
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_VIEWPORT, 18);
assert(firstViewport == 0 && viewportCount == 1);
- tu6_emit_viewport(draw_cs, pViewports);
- tu_cs_sanity_check(draw_cs);
+ tu6_emit_viewport(&cs, pViewports);
}
void
const VkRect2D *pScissors)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- struct tu_cs *draw_cs = &cmd->draw_cs;
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_SCISSOR, 3);
assert(firstScissor == 0 && scissorCount == 1);
- tu6_emit_scissor(draw_cs, pScissors);
- tu_cs_sanity_check(draw_cs);
+ tu6_emit_scissor(&cs, pScissors);
}
void
tu_CmdSetLineWidth(VkCommandBuffer commandBuffer, float lineWidth)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_LINE_WIDTH, 2);
- cmd->state.dynamic.line_width = lineWidth;
+ cmd->state.dynamic_gras_su_cntl &= ~A6XX_GRAS_SU_CNTL_LINEHALFWIDTH__MASK;
+ cmd->state.dynamic_gras_su_cntl |= A6XX_GRAS_SU_CNTL_LINEHALFWIDTH(lineWidth / 2.0f);
- /* line width depends on VkPipelineRasterizationStateCreateInfo */
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH;
+ tu_cs_emit_regs(&cs, A6XX_GRAS_SU_CNTL(.dword = cmd->state.dynamic_gras_su_cntl));
}
void
float depthBiasSlopeFactor)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- struct tu_cs *draw_cs = &cmd->draw_cs;
-
- tu6_emit_depth_bias(draw_cs, depthBiasConstantFactor, depthBiasClamp,
- depthBiasSlopeFactor);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_DEPTH_BIAS, 4);
- tu_cs_sanity_check(draw_cs);
+ tu6_emit_depth_bias(&cs, depthBiasConstantFactor, depthBiasClamp, depthBiasSlopeFactor);
}
void
const float blendConstants[4])
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- struct tu_cs *draw_cs = &cmd->draw_cs;
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_BLEND_CONSTANTS, 5);
- tu6_emit_blend_constants(draw_cs, blendConstants);
-
- tu_cs_sanity_check(draw_cs);
+ tu_cs_emit_pkt4(&cs, REG_A6XX_RB_BLEND_RED_F32, 4);
+ tu_cs_emit_array(&cs, (const uint32_t *) blendConstants, 4);
}
void
float minDepthBounds,
float maxDepthBounds)
{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_DEPTH_BOUNDS, 3);
+
+ tu_cs_emit_regs(&cs,
+ A6XX_RB_Z_BOUNDS_MIN(minDepthBounds),
+ A6XX_RB_Z_BOUNDS_MAX(maxDepthBounds));
+}
+
+static void
+update_stencil_mask(uint32_t *value, VkStencilFaceFlags face, uint32_t mask)
+{
+ if (face & VK_STENCIL_FACE_FRONT_BIT)
+ *value = (*value & 0xff00) | (mask & 0xff);
+ if (face & VK_STENCIL_FACE_BACK_BIT)
+ *value = (*value & 0xff) | (mask & 0xff) << 8;
}
void
uint32_t compareMask)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK, 2);
- if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
- cmd->state.dynamic.stencil_compare_mask.front = compareMask;
- if (faceMask & VK_STENCIL_FACE_BACK_BIT)
- cmd->state.dynamic.stencil_compare_mask.back = compareMask;
+ update_stencil_mask(&cmd->state.dynamic_stencil_mask, faceMask, compareMask);
- /* the front/back compare masks must be updated together */
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK;
+ tu_cs_emit_regs(&cs, A6XX_RB_STENCILMASK(.dword = cmd->state.dynamic_stencil_mask));
}
void
uint32_t writeMask)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_WRITE_MASK, 2);
- if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
- cmd->state.dynamic.stencil_write_mask.front = writeMask;
- if (faceMask & VK_STENCIL_FACE_BACK_BIT)
- cmd->state.dynamic.stencil_write_mask.back = writeMask;
+ update_stencil_mask(&cmd->state.dynamic_stencil_wrmask, faceMask, writeMask);
- /* the front/back write masks must be updated together */
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK;
+ tu_cs_emit_regs(&cs, A6XX_RB_STENCILWRMASK(.dword = cmd->state.dynamic_stencil_wrmask));
}
void
uint32_t reference)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, VK_DYNAMIC_STATE_STENCIL_REFERENCE, 2);
- if (faceMask & VK_STENCIL_FACE_FRONT_BIT)
- cmd->state.dynamic.stencil_reference.front = reference;
- if (faceMask & VK_STENCIL_FACE_BACK_BIT)
- cmd->state.dynamic.stencil_reference.back = reference;
+ update_stencil_mask(&cmd->state.dynamic_stencil_ref, faceMask, reference);
- /* the front/back references must be updated together */
- cmd->state.dirty |= TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE;
+ tu_cs_emit_regs(&cs, A6XX_RB_STENCILREF(.dword = cmd->state.dynamic_stencil_ref));
}
+void
+tu_CmdSetSampleLocationsEXT(VkCommandBuffer commandBuffer,
+ const VkSampleLocationsInfoEXT* pSampleLocationsInfo)
+{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs cs = tu_cmd_dynamic_state(cmd, TU_DYNAMIC_STATE_SAMPLE_LOCATIONS, 9);
+
+ assert(pSampleLocationsInfo);
+
+ tu6_emit_sample_locations(&cs, pSampleLocationsInfo);
+}
+
+static void
+tu_flush_for_access(struct tu_cache_state *cache,
+ enum tu_cmd_access_mask src_mask,
+ enum tu_cmd_access_mask dst_mask)
+{
+ enum tu_cmd_flush_bits flush_bits = 0;
+
+ if (src_mask & TU_ACCESS_HOST_WRITE) {
+ /* Host writes are always visible to CP, so only invalidate GPU caches */
+ cache->pending_flush_bits |= TU_CMD_FLAG_GPU_INVALIDATE;
+ }
+
+ if (src_mask & TU_ACCESS_SYSMEM_WRITE) {
+ /* Invalidate CP and 2D engine (make it do WFI + WFM if necessary) as
+ * well.
+ */
+ cache->pending_flush_bits |= TU_CMD_FLAG_ALL_INVALIDATE;
+ }
+
+ if (src_mask & TU_ACCESS_CP_WRITE) {
+ /* Flush the CP write queue. However a WFI shouldn't be necessary as
+ * WAIT_MEM_WRITES should cover it.
+ */
+ cache->pending_flush_bits |=
+ TU_CMD_FLAG_WAIT_MEM_WRITES |
+ TU_CMD_FLAG_GPU_INVALIDATE |
+ TU_CMD_FLAG_WAIT_FOR_ME;
+ }
+
+#define SRC_FLUSH(domain, flush, invalidate) \
+ if (src_mask & TU_ACCESS_##domain##_WRITE) { \
+ cache->pending_flush_bits |= TU_CMD_FLAG_##flush | \
+ (TU_CMD_FLAG_ALL_INVALIDATE & ~TU_CMD_FLAG_##invalidate); \
+ }
+
+ SRC_FLUSH(UCHE, CACHE_FLUSH, CACHE_INVALIDATE)
+ SRC_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
+ SRC_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+
+#undef SRC_FLUSH
+
+#define SRC_INCOHERENT_FLUSH(domain, flush, invalidate) \
+ if (src_mask & TU_ACCESS_##domain##_INCOHERENT_WRITE) { \
+ flush_bits |= TU_CMD_FLAG_##flush; \
+ cache->pending_flush_bits |= \
+ (TU_CMD_FLAG_ALL_INVALIDATE & ~TU_CMD_FLAG_##invalidate); \
+ }
+
+ SRC_INCOHERENT_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
+ SRC_INCOHERENT_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+
+#undef SRC_INCOHERENT_FLUSH
+
+ /* Treat host & sysmem write accesses the same, since the kernel implicitly
+ * drains the queue before signalling completion to the host.
+ */
+ if (dst_mask & (TU_ACCESS_SYSMEM_READ | TU_ACCESS_SYSMEM_WRITE |
+ TU_ACCESS_HOST_READ | TU_ACCESS_HOST_WRITE)) {
+ flush_bits |= cache->pending_flush_bits & TU_CMD_FLAG_ALL_FLUSH;
+ }
+
+#define DST_FLUSH(domain, flush, invalidate) \
+ if (dst_mask & (TU_ACCESS_##domain##_READ | \
+ TU_ACCESS_##domain##_WRITE)) { \
+ flush_bits |= cache->pending_flush_bits & \
+ (TU_CMD_FLAG_##invalidate | \
+ (TU_CMD_FLAG_ALL_FLUSH & ~TU_CMD_FLAG_##flush)); \
+ }
+
+ DST_FLUSH(UCHE, CACHE_FLUSH, CACHE_INVALIDATE)
+ DST_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
+ DST_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+
+#undef DST_FLUSH
+
+#define DST_INCOHERENT_FLUSH(domain, flush, invalidate) \
+ if (dst_mask & (TU_ACCESS_##domain##_READ | \
+ TU_ACCESS_##domain##_WRITE)) { \
+ flush_bits |= TU_CMD_FLAG_##invalidate | \
+ (cache->pending_flush_bits & \
+ (TU_CMD_FLAG_ALL_FLUSH & ~TU_CMD_FLAG_##flush)); \
+ }
+
+ DST_INCOHERENT_FLUSH(CCU_COLOR, CCU_FLUSH_COLOR, CCU_INVALIDATE_COLOR)
+ DST_INCOHERENT_FLUSH(CCU_DEPTH, CCU_FLUSH_DEPTH, CCU_INVALIDATE_DEPTH)
+
+#undef DST_INCOHERENT_FLUSH
+
+ if (dst_mask & TU_ACCESS_WFI_READ) {
+ flush_bits |= cache->pending_flush_bits &
+ (TU_CMD_FLAG_ALL_FLUSH | TU_CMD_FLAG_WAIT_FOR_IDLE);
+ }
+
+ if (dst_mask & TU_ACCESS_WFM_READ) {
+ flush_bits |= cache->pending_flush_bits &
+ (TU_CMD_FLAG_ALL_FLUSH | TU_CMD_FLAG_WAIT_FOR_ME);
+ }
+
+ cache->flush_bits |= flush_bits;
+ cache->pending_flush_bits &= ~flush_bits;
+}
+
+static enum tu_cmd_access_mask
+vk2tu_access(VkAccessFlags flags, bool gmem)
+{
+ enum tu_cmd_access_mask mask = 0;
+
+ /* If the GPU writes a buffer that is then read by an indirect draw
+ * command, we theoretically need to emit a WFI to wait for any cache
+ * flushes, and then a WAIT_FOR_ME to wait on the CP for the WFI to
+ * complete. Waiting for the WFI to complete is performed as part of the
+ * draw by the firmware, so we just need to execute the WFI.
+ *
+ * Transform feedback counters are read via CP_MEM_TO_REG, which implicitly
+ * does CP_WAIT_FOR_ME, but we still need a WFI if the GPU writes it.
+ */
+ if (flags &
+ (VK_ACCESS_INDIRECT_COMMAND_READ_BIT |
+ VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT |
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ mask |= TU_ACCESS_WFI_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_INDIRECT_COMMAND_READ_BIT | /* Read performed by CP */
+ VK_ACCESS_CONDITIONAL_RENDERING_READ_BIT_EXT | /* Read performed by CP */
+ VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_READ_BIT_EXT | /* Read performed by CP */
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ mask |= TU_ACCESS_SYSMEM_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_TRANSFORM_FEEDBACK_COUNTER_WRITE_BIT_EXT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ mask |= TU_ACCESS_CP_WRITE;
+ }
+
+ if (flags &
+ (VK_ACCESS_HOST_READ_BIT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ mask |= TU_ACCESS_HOST_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_HOST_WRITE_BIT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ mask |= TU_ACCESS_HOST_WRITE;
+ }
+
+ if (flags &
+ (VK_ACCESS_INDEX_READ_BIT | /* Read performed by PC, I think */
+ VK_ACCESS_VERTEX_ATTRIBUTE_READ_BIT | /* Read performed by VFD */
+ VK_ACCESS_UNIFORM_READ_BIT | /* Read performed by SP */
+ /* TODO: Is there a no-cache bit for textures so that we can ignore
+ * these?
+ */
+ VK_ACCESS_INPUT_ATTACHMENT_READ_BIT | /* Read performed by TP */
+ VK_ACCESS_SHADER_READ_BIT | /* Read perfomed by SP/TP */
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ mask |= TU_ACCESS_UCHE_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_SHADER_WRITE_BIT | /* Write performed by SP */
+ VK_ACCESS_TRANSFORM_FEEDBACK_WRITE_BIT_EXT | /* Write performed by VPC */
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ mask |= TU_ACCESS_UCHE_WRITE;
+ }
+
+ /* When using GMEM, the CCU is always flushed automatically to GMEM, and
+ * then GMEM is flushed to sysmem. Furthermore, we already had to flush any
+ * previous writes in sysmem mode when transitioning to GMEM. Therefore we
+ * can ignore CCU and pretend that color attachments and transfers use
+ * sysmem directly.
+ */
+
+ if (flags &
+ (VK_ACCESS_COLOR_ATTACHMENT_READ_BIT |
+ VK_ACCESS_COLOR_ATTACHMENT_READ_NONCOHERENT_BIT_EXT |
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ if (gmem)
+ mask |= TU_ACCESS_SYSMEM_READ;
+ else
+ mask |= TU_ACCESS_CCU_COLOR_INCOHERENT_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_READ_BIT |
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ if (gmem)
+ mask |= TU_ACCESS_SYSMEM_READ;
+ else
+ mask |= TU_ACCESS_CCU_DEPTH_INCOHERENT_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ if (gmem) {
+ mask |= TU_ACCESS_SYSMEM_WRITE;
+ } else {
+ mask |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
+ }
+ }
+
+ if (flags &
+ (VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ if (gmem) {
+ mask |= TU_ACCESS_SYSMEM_WRITE;
+ } else {
+ mask |= TU_ACCESS_CCU_DEPTH_INCOHERENT_WRITE;
+ }
+ }
+
+ /* When the dst access is a transfer read/write, it seems we sometimes need
+ * to insert a WFI after any flushes, to guarantee that the flushes finish
+ * before the 2D engine starts. However the opposite (i.e. a WFI after
+ * CP_BLIT and before any subsequent flush) does not seem to be needed, and
+ * the blob doesn't emit such a WFI.
+ */
+
+ if (flags &
+ (VK_ACCESS_TRANSFER_WRITE_BIT |
+ VK_ACCESS_MEMORY_WRITE_BIT)) {
+ if (gmem) {
+ mask |= TU_ACCESS_SYSMEM_WRITE;
+ } else {
+ mask |= TU_ACCESS_CCU_COLOR_WRITE;
+ }
+ mask |= TU_ACCESS_WFI_READ;
+ }
+
+ if (flags &
+ (VK_ACCESS_TRANSFER_READ_BIT | /* Access performed by TP */
+ VK_ACCESS_MEMORY_READ_BIT)) {
+ mask |= TU_ACCESS_UCHE_READ | TU_ACCESS_WFI_READ;
+ }
+
+ return mask;
+}
+
+
void
tu_CmdExecuteCommands(VkCommandBuffer commandBuffer,
uint32_t commandBufferCount,
assert(commandBufferCount > 0);
+ /* Emit any pending flushes. */
+ if (cmd->state.pass) {
+ tu_flush_all_pending(&cmd->state.renderpass_cache);
+ tu_emit_cache_flush_renderpass(cmd, &cmd->draw_cs);
+ } else {
+ tu_flush_all_pending(&cmd->state.cache);
+ tu_emit_cache_flush(cmd, &cmd->cs);
+ }
+
for (uint32_t i = 0; i < commandBufferCount; i++) {
TU_FROM_HANDLE(tu_cmd_buffer, secondary, pCmdBuffers[i]);
cmd->record_result = result;
break;
}
+
+ if (secondary->has_tess)
+ cmd->has_tess = true;
} else {
assert(tu_cs_is_empty(&secondary->draw_cs));
assert(tu_cs_is_empty(&secondary->draw_epilogue_cs));
MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
}
- tu_cs_emit_call(&cmd->cs, &secondary->cs);
+ tu_cs_add_entries(&cmd->cs, &secondary->cs);
}
+
+ cmd->state.index_size = secondary->state.index_size; /* for restart index update */
}
cmd->state.dirty = ~0u; /* TODO: set dirty only what needs to be */
+
+ /* After executing secondary command buffers, there may have been arbitrary
+ * flushes executed, so when we encounter a pipeline barrier with a
+ * srcMask, we have to assume that we need to invalidate. Therefore we need
+ * to re-initialize the cache with all pending invalidate bits set.
+ */
+ if (cmd->state.pass) {
+ tu_cache_init(&cmd->state.renderpass_cache);
+ } else {
+ tu_cache_init(&cmd->state.cache);
+ }
}
VkResult
TU_FROM_HANDLE(tu_device, device, _device);
struct tu_cmd_pool *pool;
- pool = vk_alloc2(&device->alloc, pAllocator, sizeof(*pool), 8,
- VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+ pool = vk_object_alloc(&device->vk, pAllocator, sizeof(*pool),
+ VK_OBJECT_TYPE_COMMAND_POOL);
if (pool == NULL)
return vk_error(device->instance, VK_ERROR_OUT_OF_HOST_MEMORY);
if (pAllocator)
pool->alloc = *pAllocator;
else
- pool->alloc = device->alloc;
+ pool->alloc = device->vk.alloc;
list_inithead(&pool->cmd_buffers);
list_inithead(&pool->free_cmd_buffers);
tu_cmd_buffer_destroy(cmd_buffer);
}
- vk_free2(&device->alloc, pAllocator, pool);
+ vk_object_free(&device->vk, pAllocator, pool);
}
VkResult
}
}
+static void
+tu_subpass_barrier(struct tu_cmd_buffer *cmd_buffer,
+ const struct tu_subpass_barrier *barrier,
+ bool external)
+{
+ /* Note: we don't know until the end of the subpass whether we'll use
+ * sysmem, so assume sysmem here to be safe.
+ */
+ struct tu_cache_state *cache =
+ external ? &cmd_buffer->state.cache : &cmd_buffer->state.renderpass_cache;
+ enum tu_cmd_access_mask src_flags =
+ vk2tu_access(barrier->src_access_mask, false);
+ enum tu_cmd_access_mask dst_flags =
+ vk2tu_access(barrier->dst_access_mask, false);
+
+ if (barrier->incoherent_ccu_color)
+ src_flags |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
+ if (barrier->incoherent_ccu_depth)
+ src_flags |= TU_ACCESS_CCU_DEPTH_INCOHERENT_WRITE;
+
+ tu_flush_for_access(cache, src_flags, dst_flags);
+}
+
void
tu_CmdBeginRenderPass(VkCommandBuffer commandBuffer,
const VkRenderPassBeginInfo *pRenderPassBegin,
cmd->state.pass = pass;
cmd->state.subpass = pass->subpasses;
cmd->state.framebuffer = fb;
+ cmd->state.render_area = pRenderPassBegin->renderArea;
- tu_cmd_update_tiling_config(cmd, &pRenderPassBegin->renderArea);
tu_cmd_prepare_tile_store_ib(cmd);
- tu_emit_load_clear(cmd, pRenderPassBegin);
+ /* Note: because this is external, any flushes will happen before draw_cs
+ * gets called. However deferred flushes could have to happen later as part
+ * of the subpass.
+ */
+ tu_subpass_barrier(cmd, &pass->subpasses[0].start_barrier, true);
+ cmd->state.renderpass_cache.pending_flush_bits =
+ cmd->state.cache.pending_flush_bits;
+ cmd->state.renderpass_cache.flush_bits = 0;
+
+ tu_emit_renderpass_begin(cmd, pRenderPassBegin);
tu6_emit_zs(cmd, cmd->state.subpass, &cmd->draw_cs);
tu6_emit_mrt(cmd, cmd->state.subpass, &cmd->draw_cs);
- tu6_emit_msaa(cmd, cmd->state.subpass, &cmd->draw_cs);
+ tu6_emit_msaa(&cmd->draw_cs, cmd->state.subpass->samples);
tu6_emit_render_cntl(cmd, cmd->state.subpass, &cmd->draw_cs, false);
- /* note: use_hw_binning only checks tiling config */
- if (use_hw_binning(cmd))
- cmd->use_vsc_data = true;
+ tu_set_input_attachments(cmd, cmd->state.subpass);
for (uint32_t i = 0; i < fb->attachment_count; ++i) {
const struct tu_image_view *iview = fb->attachments[i].attachment;
tu_bo_list_add(&cmd->bo_list, iview->image->bo,
MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
}
+
+ cmd->state.dirty |= TU_CMD_DIRTY_DRAW_STATE;
}
void
struct tu_cs *cs = &cmd->draw_cs;
const struct tu_subpass *subpass = cmd->state.subpass++;
- /* TODO:
- * if msaa samples change between subpasses,
- * attachment store is broken for some attachments
- */
+
+ tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_GMEM);
+
if (subpass->resolve_attachments) {
tu6_emit_blit_scissor(cmd, cs, true);
+
for (unsigned i = 0; i < subpass->color_count; i++) {
uint32_t a = subpass->resolve_attachments[i].attachment;
- if (a != VK_ATTACHMENT_UNUSED) {
- tu6_emit_resolve(cmd, cs, a,
- subpass->color_attachments[i].attachment);
- }
+ if (a == VK_ATTACHMENT_UNUSED)
+ continue;
+
+ tu_store_gmem_attachment(cmd, cs, a,
+ subpass->color_attachments[i].attachment);
+
+ if (pass->attachments[a].gmem_offset < 0)
+ continue;
+
+ /* TODO:
+ * check if the resolved attachment is needed by later subpasses,
+ * if it is, should be doing a GMEM->GMEM resolve instead of GMEM->MEM->GMEM..
+ */
+ tu_finishme("missing GMEM->GMEM resolve path\n");
+ tu_load_gmem_attachment(cmd, cs, a, true);
}
}
- /* invalidate because reading input attachments will cache GMEM and
- * the cache isn''t updated when GMEM is written
- * TODO: is there a no-cache bit for textures?
- */
- if (cmd->state.subpass->input_count)
- tu6_emit_event_write(cmd, cs, CACHE_INVALIDATE, false);
+ tu_cond_exec_end(cs);
+
+ tu_cond_exec_start(cs, CP_COND_EXEC_0_RENDER_MODE_SYSMEM);
+
+ tu6_emit_sysmem_resolves(cmd, cs, subpass);
+
+ tu_cond_exec_end(cs);
+
+ /* Handle dependencies for the next subpass */
+ tu_subpass_barrier(cmd, &cmd->state.subpass->start_barrier, false);
/* emit mrt/zs/msaa/ubwc state for the subpass that is starting */
tu6_emit_zs(cmd, cmd->state.subpass, cs);
tu6_emit_mrt(cmd, cmd->state.subpass, cs);
- tu6_emit_msaa(cmd, cmd->state.subpass, cs);
+ tu6_emit_msaa(cs, cmd->state.subpass->samples);
tu6_emit_render_cntl(cmd, cmd->state.subpass, cs, false);
- /* Emit flushes so that input attachments will read the correct value. This
- * is for sysmem only, although it shouldn't do much harm on gmem.
- */
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_COLOR_TS, true);
- tu6_emit_event_write(cmd, cs, PC_CCU_FLUSH_DEPTH_TS, true);
-
- /* TODO:
- * since we don't know how to do GMEM->GMEM resolve,
- * resolve attachments are resolved to memory then loaded to GMEM again if needed
- */
- if (subpass->resolve_attachments) {
- for (unsigned i = 0; i < subpass->color_count; i++) {
- uint32_t a = subpass->resolve_attachments[i].attachment;
- if (a != VK_ATTACHMENT_UNUSED && pass->attachments[a].gmem_offset >= 0) {
- tu_finishme("missing GMEM->GMEM resolve, performance will suffer\n");
- tu6_emit_predicated_blit(cmd, cs, a, a, false);
- }
- }
- }
+ tu_set_input_attachments(cmd, cmd->state.subpass);
}
void
tu_CmdNextSubpass(commandBuffer, pSubpassBeginInfo->contents);
}
-struct tu_draw_info
-{
- /**
- * Number of vertices.
- */
- uint32_t count;
-
- /**
- * Index of the first vertex.
- */
- int32_t vertex_offset;
-
- /**
- * First instance id.
- */
- uint32_t first_instance;
-
- /**
- * Number of instances.
- */
- uint32_t instance_count;
-
- /**
- * First index (indexed draws only).
- */
- uint32_t first_index;
-
- /**
- * Whether it's an indexed draw.
- */
- bool indexed;
-
- /**
- * Indirect draw parameters resource.
- */
- struct tu_buffer *indirect;
- uint64_t indirect_offset;
- uint32_t stride;
-
- /**
- * Draw count parameters resource.
- */
- struct tu_buffer *count_buffer;
- uint64_t count_buffer_offset;
-
- /**
- * Stream output parameters resource.
- */
- struct tu_buffer *streamout_buffer;
- uint64_t streamout_buffer_offset;
-};
-
-#define ENABLE_ALL (CP_SET_DRAW_STATE__0_BINNING | CP_SET_DRAW_STATE__0_GMEM | CP_SET_DRAW_STATE__0_SYSMEM)
-#define ENABLE_DRAW (CP_SET_DRAW_STATE__0_GMEM | CP_SET_DRAW_STATE__0_SYSMEM)
-
-enum tu_draw_state_group_id
-{
- TU_DRAW_STATE_PROGRAM,
- TU_DRAW_STATE_PROGRAM_BINNING,
- TU_DRAW_STATE_VI,
- TU_DRAW_STATE_VI_BINNING,
- TU_DRAW_STATE_VP,
- TU_DRAW_STATE_RAST,
- TU_DRAW_STATE_DS,
- TU_DRAW_STATE_BLEND,
- TU_DRAW_STATE_VS_CONST,
- TU_DRAW_STATE_FS_CONST,
- TU_DRAW_STATE_VS_TEX,
- TU_DRAW_STATE_FS_TEX_SYSMEM,
- TU_DRAW_STATE_FS_TEX_GMEM,
- TU_DRAW_STATE_FS_IBO,
- TU_DRAW_STATE_VS_PARAMS,
-
- TU_DRAW_STATE_COUNT,
-};
-
-struct tu_draw_state_group
-{
- enum tu_draw_state_group_id id;
- uint32_t enable_mask;
- struct tu_cs_entry ib;
-};
-
-const static void *
-sampler_ptr(struct tu_descriptor_state *descriptors_state,
- const struct tu_descriptor_map *map, unsigned i,
- unsigned array_index)
-{
- assert(descriptors_state->valid & (1 << map->set[i]));
-
- struct tu_descriptor_set *set = descriptors_state->sets[map->set[i]];
- assert(map->binding[i] < set->layout->binding_count);
-
- const struct tu_descriptor_set_binding_layout *layout =
- &set->layout->binding[map->binding[i]];
-
- if (layout->immutable_samplers_offset) {
- const uint32_t *immutable_samplers =
- tu_immutable_samplers(set->layout, layout);
-
- return &immutable_samplers[array_index * A6XX_TEX_SAMP_DWORDS];
- }
-
- switch (layout->type) {
- case VK_DESCRIPTOR_TYPE_SAMPLER:
- return &set->mapped_ptr[layout->offset / 4];
- case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
- return &set->mapped_ptr[layout->offset / 4 + A6XX_TEX_CONST_DWORDS +
- array_index * (A6XX_TEX_CONST_DWORDS + A6XX_TEX_SAMP_DWORDS)];
- default:
- unreachable("unimplemented descriptor type");
- break;
- }
-}
-
-static void
-write_tex_const(struct tu_cmd_buffer *cmd,
- uint32_t *dst,
- struct tu_descriptor_state *descriptors_state,
- const struct tu_descriptor_map *map,
- unsigned i, unsigned array_index, bool is_sysmem)
-{
- assert(descriptors_state->valid & (1 << map->set[i]));
-
- struct tu_descriptor_set *set = descriptors_state->sets[map->set[i]];
- assert(map->binding[i] < set->layout->binding_count);
-
- const struct tu_descriptor_set_binding_layout *layout =
- &set->layout->binding[map->binding[i]];
-
- switch (layout->type) {
- case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
- case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
- case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
- case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
- memcpy(dst, &set->mapped_ptr[layout->offset / 4 +
- array_index * A6XX_TEX_CONST_DWORDS],
- A6XX_TEX_CONST_DWORDS * 4);
- break;
- case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
- memcpy(dst, &set->mapped_ptr[layout->offset / 4 +
- array_index *
- (A6XX_TEX_CONST_DWORDS +
- A6XX_TEX_SAMP_DWORDS)],
- A6XX_TEX_CONST_DWORDS * 4);
- break;
- default:
- unreachable("unimplemented descriptor type");
- break;
- }
-
- if (layout->type == VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT && !is_sysmem) {
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
- uint32_t a = cmd->state.subpass->input_attachments[map->value[i] +
- array_index].attachment;
- const struct tu_render_pass_attachment *att = &cmd->state.pass->attachments[a];
-
- assert(att->gmem_offset >= 0);
-
- dst[0] &= ~(A6XX_TEX_CONST_0_SWAP__MASK | A6XX_TEX_CONST_0_TILE_MODE__MASK);
- dst[0] |= A6XX_TEX_CONST_0_TILE_MODE(TILE6_2);
- dst[2] &= ~(A6XX_TEX_CONST_2_TYPE__MASK | A6XX_TEX_CONST_2_PITCH__MASK);
- dst[2] |=
- A6XX_TEX_CONST_2_TYPE(A6XX_TEX_2D) |
- A6XX_TEX_CONST_2_PITCH(tiling->tile0.extent.width * att->cpp);
- dst[3] = 0;
- dst[4] = cmd->device->physical_device->gmem_base + att->gmem_offset;
- dst[5] = A6XX_TEX_CONST_5_DEPTH(1);
- for (unsigned i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
- dst[i] = 0;
-
- if (cmd->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY)
- tu_finishme("patch input attachment pitch for secondary cmd buffer");
- }
-}
-
-static void
-write_image_ibo(struct tu_cmd_buffer *cmd,
- uint32_t *dst,
- struct tu_descriptor_state *descriptors_state,
- const struct tu_descriptor_map *map,
- unsigned i, unsigned array_index)
-{
- assert(descriptors_state->valid & (1 << map->set[i]));
-
- struct tu_descriptor_set *set = descriptors_state->sets[map->set[i]];
- assert(map->binding[i] < set->layout->binding_count);
-
- const struct tu_descriptor_set_binding_layout *layout =
- &set->layout->binding[map->binding[i]];
-
- assert(layout->type == VK_DESCRIPTOR_TYPE_STORAGE_IMAGE);
-
- memcpy(dst, &set->mapped_ptr[layout->offset / 4 +
- (array_index * 2 + 1) * A6XX_TEX_CONST_DWORDS],
- A6XX_TEX_CONST_DWORDS * 4);
-}
-
-static uint64_t
-buffer_ptr(struct tu_descriptor_state *descriptors_state,
- const struct tu_descriptor_map *map,
- unsigned i, unsigned array_index)
-{
- assert(descriptors_state->valid & (1 << map->set[i]));
-
- struct tu_descriptor_set *set = descriptors_state->sets[map->set[i]];
- assert(map->binding[i] < set->layout->binding_count);
-
- const struct tu_descriptor_set_binding_layout *layout =
- &set->layout->binding[map->binding[i]];
-
- switch (layout->type) {
- case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
- case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC:
- return descriptors_state->dynamic_buffers[layout->dynamic_offset_offset +
- array_index];
- case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
- case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
- return (uint64_t) set->mapped_ptr[layout->offset / 4 + array_index * 2 + 1] << 32 |
- set->mapped_ptr[layout->offset / 4 + array_index * 2];
- default:
- unreachable("unimplemented descriptor type");
- break;
- }
-}
-
-static inline uint32_t
-tu6_stage2opcode(gl_shader_stage type)
-{
- switch (type) {
- case MESA_SHADER_VERTEX:
- case MESA_SHADER_TESS_CTRL:
- case MESA_SHADER_TESS_EVAL:
- case MESA_SHADER_GEOMETRY:
- return CP_LOAD_STATE6_GEOM;
- case MESA_SHADER_FRAGMENT:
- case MESA_SHADER_COMPUTE:
- case MESA_SHADER_KERNEL:
- return CP_LOAD_STATE6_FRAG;
- default:
- unreachable("bad shader type");
- }
-}
-
-static inline enum a6xx_state_block
-tu6_stage2shadersb(gl_shader_stage type)
-{
- switch (type) {
- case MESA_SHADER_VERTEX:
- return SB6_VS_SHADER;
- case MESA_SHADER_FRAGMENT:
- return SB6_FS_SHADER;
- case MESA_SHADER_COMPUTE:
- case MESA_SHADER_KERNEL:
- return SB6_CS_SHADER;
- default:
- unreachable("bad shader type");
- return ~0;
- }
-}
-
static void
tu6_emit_user_consts(struct tu_cs *cs, const struct tu_pipeline *pipeline,
struct tu_descriptor_state *descriptors_state,
{
const struct tu_program_descriptor_linkage *link =
&pipeline->program.link[type];
- const struct ir3_ubo_analysis_state *state = &link->ubo_state;
-
- for (uint32_t i = 0; i < ARRAY_SIZE(state->range); i++) {
- if (state->range[i].start < state->range[i].end) {
- uint32_t size = state->range[i].end - state->range[i].start;
- uint32_t offset = state->range[i].start;
-
- /* and even if the start of the const buffer is before
- * first_immediate, the end may not be:
- */
- size = MIN2(size, (16 * link->constlen) - state->range[i].offset);
-
- if (size == 0)
- continue;
-
- /* things should be aligned to vec4: */
- debug_assert((state->range[i].offset % 16) == 0);
- debug_assert((size % 16) == 0);
- debug_assert((offset % 16) == 0);
-
- if (i == 0) {
- /* push constants */
- tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + (size / 4));
- tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(state->range[i].offset / 16) |
- CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
- CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
- CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
- CP_LOAD_STATE6_0_NUM_UNIT(size / 16));
- tu_cs_emit(cs, 0);
- tu_cs_emit(cs, 0);
- for (unsigned i = 0; i < size / 4; i++)
- tu_cs_emit(cs, push_constants[i + offset / 4]);
- continue;
- }
+ const struct ir3_ubo_analysis_state *state = &link->const_state.ubo_state;
- /* Look through the UBO map to find our UBO index, and get the VA for
- * that UBO.
- */
- uint64_t va = 0;
- uint32_t ubo_idx = i - 1;
- uint32_t ubo_map_base = 0;
- for (int j = 0; j < link->ubo_map.num; j++) {
- if (ubo_idx >= ubo_map_base &&
- ubo_idx < ubo_map_base + link->ubo_map.array_size[j]) {
- va = buffer_ptr(descriptors_state, &link->ubo_map, j,
- ubo_idx - ubo_map_base);
- break;
- }
- ubo_map_base += link->ubo_map.array_size[j];
- }
- assert(va);
-
- tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3);
- tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(state->range[i].offset / 16) |
- CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
- CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
- CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
- CP_LOAD_STATE6_0_NUM_UNIT(size / 16));
- tu_cs_emit_qw(cs, va + offset);
- }
+ if (link->push_consts.count > 0) {
+ unsigned num_units = link->push_consts.count;
+ unsigned offset = link->push_consts.lo;
+ tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + num_units * 4);
+ tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(offset) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
+ CP_LOAD_STATE6_0_NUM_UNIT(num_units));
+ tu_cs_emit(cs, 0);
+ tu_cs_emit(cs, 0);
+ for (unsigned i = 0; i < num_units * 4; i++)
+ tu_cs_emit(cs, push_constants[i + offset * 4]);
}
-}
-
-static void
-tu6_emit_ubos(struct tu_cs *cs, const struct tu_pipeline *pipeline,
- struct tu_descriptor_state *descriptors_state,
- gl_shader_stage type)
-{
- const struct tu_program_descriptor_linkage *link =
- &pipeline->program.link[type];
- uint32_t num = MIN2(link->ubo_map.num_desc, link->const_state.num_ubos);
- uint32_t anum = align(num, 2);
+ for (uint32_t i = 0; i < state->num_enabled; i++) {
+ uint32_t size = state->range[i].end - state->range[i].start;
+ uint32_t offset = state->range[i].start;
- if (!num)
- return;
+ /* and even if the start of the const buffer is before
+ * first_immediate, the end may not be:
+ */
+ size = MIN2(size, (16 * link->constlen) - state->range[i].offset);
- tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + (2 * anum));
- tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(link->const_state.offsets.ubo) |
- CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
- CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
- CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
- CP_LOAD_STATE6_0_NUM_UNIT(anum/2));
- tu_cs_emit(cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
- tu_cs_emit(cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
-
- unsigned emitted = 0;
- for (unsigned i = 0; emitted < num && i < link->ubo_map.num; i++) {
- for (unsigned j = 0; emitted < num && j < link->ubo_map.array_size[i]; j++) {
- tu_cs_emit_qw(cs, buffer_ptr(descriptors_state, &link->ubo_map, i, j));
- emitted++;
- }
- }
+ if (size == 0)
+ continue;
- for (; emitted < anum; emitted++) {
- tu_cs_emit(cs, 0xffffffff);
- tu_cs_emit(cs, 0xffffffff);
- }
-}
+ /* things should be aligned to vec4: */
+ debug_assert((state->range[i].offset % 16) == 0);
+ debug_assert((size % 16) == 0);
+ debug_assert((offset % 16) == 0);
-static struct tu_cs_entry
+ /* Dig out the descriptor from the descriptor state and read the VA from
+ * it.
+ */
+ assert(state->range[i].ubo.bindless);
+ uint32_t *base = state->range[i].ubo.bindless_base == MAX_SETS ?
+ descriptors_state->dynamic_descriptors :
+ descriptors_state->sets[state->range[i].ubo.bindless_base]->mapped_ptr;
+ unsigned block = state->range[i].ubo.block;
+ uint32_t *desc = base + block * A6XX_TEX_CONST_DWORDS;
+ uint64_t va = desc[0] | ((uint64_t)(desc[1] & A6XX_UBO_1_BASE_HI__MASK) << 32);
+ assert(va);
+
+ tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3);
+ tu_cs_emit(cs, CP_LOAD_STATE6_0_DST_OFF(state->range[i].offset / 16) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(tu6_stage2shadersb(type)) |
+ CP_LOAD_STATE6_0_NUM_UNIT(size / 16));
+ tu_cs_emit_qw(cs, va + offset);
+ }
+}
+
+static struct tu_draw_state
tu6_emit_consts(struct tu_cmd_buffer *cmd,
const struct tu_pipeline *pipeline,
struct tu_descriptor_state *descriptors_state,
tu_cs_begin_sub_stream(&cmd->sub_cs, 512, &cs); /* TODO: maximum size? */
tu6_emit_user_consts(&cs, pipeline, descriptors_state, type, cmd->push_constants);
- tu6_emit_ubos(&cs, pipeline, descriptors_state, type);
-
- return tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
-}
-
-static VkResult
-tu6_emit_vs_params(struct tu_cmd_buffer *cmd,
- const struct tu_draw_info *draw,
- struct tu_cs_entry *entry)
-{
- /* TODO: fill out more than just base instance */
- const struct tu_program_descriptor_linkage *link =
- &cmd->state.pipeline->program.link[MESA_SHADER_VERTEX];
- const struct ir3_const_state *const_state = &link->const_state;
- struct tu_cs cs;
-
- if (const_state->offsets.driver_param >= link->constlen) {
- *entry = (struct tu_cs_entry) {};
- return VK_SUCCESS;
- }
-
- VkResult result = tu_cs_begin_sub_stream(&cmd->sub_cs, 8, &cs);
- if (result != VK_SUCCESS)
- return result;
-
- tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_GEOM, 3 + 4);
- tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(const_state->offsets.driver_param) |
- CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
- CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
- CP_LOAD_STATE6_0_STATE_BLOCK(SB6_VS_SHADER) |
- CP_LOAD_STATE6_0_NUM_UNIT(1));
- tu_cs_emit(&cs, 0);
- tu_cs_emit(&cs, 0);
-
- STATIC_ASSERT(IR3_DP_INSTID_BASE == 2);
-
- tu_cs_emit(&cs, 0);
- tu_cs_emit(&cs, 0);
- tu_cs_emit(&cs, draw->first_instance);
- tu_cs_emit(&cs, 0);
-
- *entry = tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
- return VK_SUCCESS;
-}
-
-static VkResult
-tu6_emit_textures(struct tu_cmd_buffer *cmd,
- const struct tu_pipeline *pipeline,
- struct tu_descriptor_state *descriptors_state,
- gl_shader_stage type,
- struct tu_cs_entry *entry,
- bool is_sysmem)
-{
- struct tu_cs *draw_state = &cmd->sub_cs;
- const struct tu_program_descriptor_linkage *link =
- &pipeline->program.link[type];
- VkResult result;
-
- if (link->texture_map.num_desc == 0 && link->sampler_map.num_desc == 0) {
- *entry = (struct tu_cs_entry) {};
- return VK_SUCCESS;
- }
-
- /* allocate and fill texture state */
- struct ts_cs_memory tex_const;
- result = tu_cs_alloc(draw_state, link->texture_map.num_desc,
- A6XX_TEX_CONST_DWORDS, &tex_const);
- if (result != VK_SUCCESS)
- return result;
-
- int tex_index = 0;
- for (unsigned i = 0; i < link->texture_map.num; i++) {
- for (int j = 0; j < link->texture_map.array_size[i]; j++) {
- write_tex_const(cmd,
- &tex_const.map[A6XX_TEX_CONST_DWORDS * tex_index++],
- descriptors_state, &link->texture_map, i, j,
- is_sysmem);
- }
- }
-
- /* allocate and fill sampler state */
- struct ts_cs_memory tex_samp = { 0 };
- if (link->sampler_map.num_desc) {
- result = tu_cs_alloc(draw_state, link->sampler_map.num_desc,
- A6XX_TEX_SAMP_DWORDS, &tex_samp);
- if (result != VK_SUCCESS)
- return result;
-
- int sampler_index = 0;
- for (unsigned i = 0; i < link->sampler_map.num; i++) {
- for (int j = 0; j < link->sampler_map.array_size[i]; j++) {
- const uint32_t *sampler = sampler_ptr(descriptors_state,
- &link->sampler_map,
- i, j);
- memcpy(&tex_samp.map[A6XX_TEX_SAMP_DWORDS * sampler_index++],
- sampler, A6XX_TEX_SAMP_DWORDS * 4);
- }
- }
- }
-
- unsigned tex_samp_reg, tex_const_reg, tex_count_reg;
- enum a6xx_state_block sb;
-
- switch (type) {
- case MESA_SHADER_VERTEX:
- sb = SB6_VS_TEX;
- tex_samp_reg = REG_A6XX_SP_VS_TEX_SAMP_LO;
- tex_const_reg = REG_A6XX_SP_VS_TEX_CONST_LO;
- tex_count_reg = REG_A6XX_SP_VS_TEX_COUNT;
- break;
- case MESA_SHADER_FRAGMENT:
- sb = SB6_FS_TEX;
- tex_samp_reg = REG_A6XX_SP_FS_TEX_SAMP_LO;
- tex_const_reg = REG_A6XX_SP_FS_TEX_CONST_LO;
- tex_count_reg = REG_A6XX_SP_FS_TEX_COUNT;
- break;
- case MESA_SHADER_COMPUTE:
- sb = SB6_CS_TEX;
- tex_samp_reg = REG_A6XX_SP_CS_TEX_SAMP_LO;
- tex_const_reg = REG_A6XX_SP_CS_TEX_CONST_LO;
- tex_count_reg = REG_A6XX_SP_CS_TEX_COUNT;
- break;
- default:
- unreachable("bad state block");
- }
-
- struct tu_cs cs;
- result = tu_cs_begin_sub_stream(draw_state, 16, &cs);
- if (result != VK_SUCCESS)
- return result;
-
- if (link->sampler_map.num_desc) {
- /* output sampler state: */
- tu_cs_emit_pkt7(&cs, tu6_stage2opcode(type), 3);
- tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
- CP_LOAD_STATE6_0_STATE_TYPE(ST6_SHADER) |
- CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
- CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
- CP_LOAD_STATE6_0_NUM_UNIT(link->sampler_map.num_desc));
- tu_cs_emit_qw(&cs, tex_samp.iova); /* SRC_ADDR_LO/HI */
-
- tu_cs_emit_pkt4(&cs, tex_samp_reg, 2);
- tu_cs_emit_qw(&cs, tex_samp.iova); /* SRC_ADDR_LO/HI */
- }
-
- /* emit texture state: */
- tu_cs_emit_pkt7(&cs, tu6_stage2opcode(type), 3);
- tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
- CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
- CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
- CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
- CP_LOAD_STATE6_0_NUM_UNIT(link->texture_map.num_desc));
- tu_cs_emit_qw(&cs, tex_const.iova); /* SRC_ADDR_LO/HI */
- tu_cs_emit_pkt4(&cs, tex_const_reg, 2);
- tu_cs_emit_qw(&cs, tex_const.iova); /* SRC_ADDR_LO/HI */
-
- tu_cs_emit_pkt4(&cs, tex_count_reg, 1);
- tu_cs_emit(&cs, link->texture_map.num_desc);
-
- *entry = tu_cs_end_sub_stream(draw_state, &cs);
- return VK_SUCCESS;
+ return tu_cs_end_draw_state(&cmd->sub_cs, &cs);
}
-static VkResult
-tu6_emit_ibo(struct tu_cmd_buffer *cmd,
- const struct tu_pipeline *pipeline,
- struct tu_descriptor_state *descriptors_state,
- gl_shader_stage type,
- struct tu_cs_entry *entry)
-{
- struct tu_cs *draw_state = &cmd->sub_cs;
- const struct tu_program_descriptor_linkage *link =
- &pipeline->program.link[type];
- VkResult result;
-
- unsigned num_desc = link->ssbo_map.num_desc + link->image_map.num_desc;
-
- if (num_desc == 0) {
- *entry = (struct tu_cs_entry) {};
- return VK_SUCCESS;
- }
-
- struct ts_cs_memory ibo_const;
- result = tu_cs_alloc(draw_state, num_desc,
- A6XX_TEX_CONST_DWORDS, &ibo_const);
- if (result != VK_SUCCESS)
- return result;
-
- int ssbo_index = 0;
- for (unsigned i = 0; i < link->ssbo_map.num; i++) {
- for (int j = 0; j < link->ssbo_map.array_size[i]; j++) {
- uint32_t *dst = &ibo_const.map[A6XX_TEX_CONST_DWORDS * ssbo_index];
-
- uint64_t va = buffer_ptr(descriptors_state, &link->ssbo_map, i, j);
- /* We don't expose robustBufferAccess, so leave the size unlimited. */
- uint32_t sz = MAX_STORAGE_BUFFER_RANGE / 4;
-
- dst[0] = A6XX_IBO_0_FMT(FMT6_32_UINT);
- dst[1] = A6XX_IBO_1_WIDTH(sz & MASK(15)) |
- A6XX_IBO_1_HEIGHT(sz >> 15);
- dst[2] = A6XX_IBO_2_UNK4 |
- A6XX_IBO_2_UNK31 |
- A6XX_IBO_2_TYPE(A6XX_TEX_1D);
- dst[3] = 0;
- dst[4] = va;
- dst[5] = va >> 32;
- for (int i = 6; i < A6XX_TEX_CONST_DWORDS; i++)
- dst[i] = 0;
-
- ssbo_index++;
- }
- }
+static struct tu_draw_state
+tu6_emit_vertex_buffers(struct tu_cmd_buffer *cmd,
+ const struct tu_pipeline *pipeline)
+{
+ struct tu_cs cs;
+ tu_cs_begin_sub_stream(&cmd->sub_cs, 4 * MAX_VBS, &cs);
- for (unsigned i = 0; i < link->image_map.num; i++) {
- for (int j = 0; j < link->image_map.array_size[i]; j++) {
- uint32_t *dst = &ibo_const.map[A6XX_TEX_CONST_DWORDS * ssbo_index];
+ int binding;
+ for_each_bit(binding, pipeline->vi.bindings_used) {
+ const struct tu_buffer *buf = cmd->state.vb.buffers[binding];
+ const VkDeviceSize offset = buf->bo_offset +
+ cmd->state.vb.offsets[binding];
- write_image_ibo(cmd, dst,
- descriptors_state, &link->image_map, i, j);
+ tu_cs_emit_regs(&cs,
+ A6XX_VFD_FETCH_BASE(binding, .bo = buf->bo, .bo_offset = offset),
+ A6XX_VFD_FETCH_SIZE(binding, buf->size - offset));
- ssbo_index++;
- }
}
- assert(ssbo_index == num_desc);
+ cmd->vertex_bindings_set = pipeline->vi.bindings_used;
- struct tu_cs cs;
- result = tu_cs_begin_sub_stream(draw_state, 7, &cs);
- if (result != VK_SUCCESS)
- return result;
+ return tu_cs_end_draw_state(&cmd->sub_cs, &cs);
+}
- uint32_t opcode, ibo_addr_reg;
- enum a6xx_state_block sb;
- enum a6xx_state_type st;
+static uint64_t
+get_tess_param_bo_size(const struct tu_pipeline *pipeline,
+ uint32_t draw_count)
+{
+ /* TODO: For indirect draws, we can't compute the BO size ahead of time.
+ * Still not sure what to do here, so just allocate a reasonably large
+ * BO and hope for the best for now. */
+ if (!draw_count)
+ draw_count = 2048;
+
+ /* the tess param BO is pipeline->tess.param_stride bytes per patch,
+ * which includes both the per-vertex outputs and per-patch outputs
+ * build_primitive_map in ir3 calculates this stride
+ */
+ uint32_t verts_per_patch = pipeline->ia.primtype - DI_PT_PATCHES0;
+ uint32_t num_patches = draw_count / verts_per_patch;
+ return num_patches * pipeline->tess.param_stride;
+}
- switch (type) {
- case MESA_SHADER_FRAGMENT:
- opcode = CP_LOAD_STATE6;
- st = ST6_SHADER;
- sb = SB6_IBO;
- ibo_addr_reg = REG_A6XX_SP_IBO_LO;
+static uint64_t
+get_tess_factor_bo_size(const struct tu_pipeline *pipeline,
+ uint32_t draw_count)
+{
+ /* TODO: For indirect draws, we can't compute the BO size ahead of time.
+ * Still not sure what to do here, so just allocate a reasonably large
+ * BO and hope for the best for now. */
+ if (!draw_count)
+ draw_count = 2048;
+
+ /* Each distinct patch gets its own tess factor output. */
+ uint32_t verts_per_patch = pipeline->ia.primtype - DI_PT_PATCHES0;
+ uint32_t num_patches = draw_count / verts_per_patch;
+ uint32_t factor_stride;
+ switch (pipeline->tess.patch_type) {
+ case IR3_TESS_ISOLINES:
+ factor_stride = 12;
+ break;
+ case IR3_TESS_TRIANGLES:
+ factor_stride = 20;
break;
- case MESA_SHADER_COMPUTE:
- opcode = CP_LOAD_STATE6_FRAG;
- st = ST6_IBO;
- sb = SB6_CS_SHADER;
- ibo_addr_reg = REG_A6XX_SP_CS_IBO_LO;
+ case IR3_TESS_QUADS:
+ factor_stride = 28;
break;
default:
- unreachable("unsupported stage for ibos");
+ unreachable("bad tessmode");
}
-
- /* emit texture state: */
- tu_cs_emit_pkt7(&cs, opcode, 3);
- tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(0) |
- CP_LOAD_STATE6_0_STATE_TYPE(st) |
- CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
- CP_LOAD_STATE6_0_STATE_BLOCK(sb) |
- CP_LOAD_STATE6_0_NUM_UNIT(num_desc));
- tu_cs_emit_qw(&cs, ibo_const.iova); /* SRC_ADDR_LO/HI */
-
- tu_cs_emit_pkt4(&cs, ibo_addr_reg, 2);
- tu_cs_emit_qw(&cs, ibo_const.iova); /* SRC_ADDR_LO/HI */
-
- *entry = tu_cs_end_sub_stream(draw_state, &cs);
- return VK_SUCCESS;
+ return factor_stride * num_patches;
}
-static void
-tu6_emit_streamout(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+static VkResult
+tu6_emit_tess_consts(struct tu_cmd_buffer *cmd,
+ uint32_t draw_count,
+ const struct tu_pipeline *pipeline,
+ struct tu_draw_state *state,
+ uint64_t *factor_iova)
{
- struct tu_streamout_state *tf = &cmd->state.pipeline->streamout;
-
- for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
- struct tu_buffer *buf = cmd->state.streamout_buf.buffers[i];
- if (!buf)
- continue;
-
- uint32_t offset;
- offset = cmd->state.streamout_buf.offsets[i];
-
- tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_BASE(i, .bo = buf->bo,
- .bo_offset = buf->bo_offset));
- tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_SIZE(i, buf->size));
-
- if (cmd->state.streamout_reset & (1 << i)) {
- offset *= tf->stride[i];
-
- tu_cs_emit_regs(cs, A6XX_VPC_SO_BUFFER_OFFSET(i, offset));
- cmd->state.streamout_reset &= ~(1 << i);
- } else {
- tu_cs_emit_pkt7(cs, CP_MEM_TO_REG, 3);
- tu_cs_emit(cs, CP_MEM_TO_REG_0_REG(REG_A6XX_VPC_SO_BUFFER_OFFSET(i)) |
- CP_MEM_TO_REG_0_SHIFT_BY_2 | CP_MEM_TO_REG_0_UNK31 |
- CP_MEM_TO_REG_0_CNT(0));
- tu_cs_emit_qw(cs, cmd->scratch_bo.iova +
- ctrl_offset(flush_base[i].offset));
- }
+ struct tu_cs cs;
+ VkResult result = tu_cs_begin_sub_stream(&cmd->sub_cs, 16, &cs);
+ if (result != VK_SUCCESS)
+ return result;
- tu_cs_emit_regs(cs, A6XX_VPC_SO_FLUSH_BASE(i, .bo = &cmd->scratch_bo,
- .bo_offset =
- ctrl_offset(flush_base[i])));
- }
+ uint64_t tess_factor_size = get_tess_factor_bo_size(pipeline, draw_count);
+ uint64_t tess_param_size = get_tess_param_bo_size(pipeline, draw_count);
+ uint64_t tess_bo_size = tess_factor_size + tess_param_size;
+ if (tess_bo_size > 0) {
+ struct tu_bo *tess_bo;
+ result = tu_get_scratch_bo(cmd->device, tess_bo_size, &tess_bo);
+ if (result != VK_SUCCESS)
+ return result;
- if (cmd->state.streamout_enabled) {
- tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 12 + (2 * tf->prog_count));
- tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
- tu_cs_emit(cs, tf->vpc_so_buf_cntl);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(0));
- tu_cs_emit(cs, tf->ncomp[0]);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(1));
- tu_cs_emit(cs, tf->ncomp[1]);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(2));
- tu_cs_emit(cs, tf->ncomp[2]);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_NCOMP(3));
- tu_cs_emit(cs, tf->ncomp[3]);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
- tu_cs_emit(cs, A6XX_VPC_SO_CNTL_ENABLE);
- for (unsigned i = 0; i < tf->prog_count; i++) {
- tu_cs_emit(cs, REG_A6XX_VPC_SO_PROG);
- tu_cs_emit(cs, tf->prog[i]);
- }
- } else {
- tu_cs_emit_pkt7(cs, CP_CONTEXT_REG_BUNCH, 4);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_CNTL);
- tu_cs_emit(cs, 0);
- tu_cs_emit(cs, REG_A6XX_VPC_SO_BUF_CNTL);
- tu_cs_emit(cs, 0);
- }
+ tu_bo_list_add(&cmd->bo_list, tess_bo,
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
+ uint64_t tess_factor_iova = tess_bo->iova;
+ uint64_t tess_param_iova = tess_factor_iova + tess_factor_size;
+
+ tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_GEOM, 3 + 4);
+ tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(pipeline->tess.hs_bo_regid) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(SB6_HS_SHADER) |
+ CP_LOAD_STATE6_0_NUM_UNIT(1));
+ tu_cs_emit(&cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
+ tu_cs_emit(&cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
+ tu_cs_emit_qw(&cs, tess_param_iova);
+ tu_cs_emit_qw(&cs, tess_factor_iova);
+
+ tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_GEOM, 3 + 4);
+ tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(pipeline->tess.ds_bo_regid) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(SB6_DS_SHADER) |
+ CP_LOAD_STATE6_0_NUM_UNIT(1));
+ tu_cs_emit(&cs, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
+ tu_cs_emit(&cs, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
+ tu_cs_emit_qw(&cs, tess_param_iova);
+ tu_cs_emit_qw(&cs, tess_factor_iova);
+
+ *factor_iova = tess_factor_iova;
+ }
+ *state = tu_cs_end_draw_state(&cmd->sub_cs, &cs);
+ return VK_SUCCESS;
}
static VkResult
-tu6_bind_draw_states(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_draw_info *draw)
+tu6_draw_common(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs,
+ bool indexed,
+ /* note: draw_count is 0 for indirect */
+ uint32_t draw_count)
{
const struct tu_pipeline *pipeline = cmd->state.pipeline;
- const struct tu_dynamic_state *dynamic = &cmd->state.dynamic;
- struct tu_draw_state_group draw_state_groups[TU_DRAW_STATE_COUNT];
- uint32_t draw_state_group_count = 0;
VkResult result;
struct tu_descriptor_state *descriptors_state =
&cmd->descriptors[VK_PIPELINE_BIND_POINT_GRAPHICS];
- /* TODO lrz */
-
- tu_cs_emit_regs(cs,
- A6XX_PC_PRIMITIVE_CNTL_0(.primitive_restart =
- pipeline->ia.primitive_restart && draw->indexed));
-
- if (cmd->state.dirty &
- (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DYNAMIC_LINE_WIDTH) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_LINE_WIDTH)) {
- tu6_emit_gras_su_cntl(cs, pipeline->rast.gras_su_cntl,
- dynamic->line_width);
- }
-
- if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_COMPARE_MASK) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_COMPARE_MASK)) {
- tu6_emit_stencil_compare_mask(cs, dynamic->stencil_compare_mask.front,
- dynamic->stencil_compare_mask.back);
- }
-
- if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_WRITE_MASK) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_WRITE_MASK)) {
- tu6_emit_stencil_write_mask(cs, dynamic->stencil_write_mask.front,
- dynamic->stencil_write_mask.back);
- }
-
- if ((cmd->state.dirty & TU_CMD_DIRTY_DYNAMIC_STENCIL_REFERENCE) &&
- (pipeline->dynamic_state.mask & TU_DYNAMIC_STENCIL_REFERENCE)) {
- tu6_emit_stencil_reference(cs, dynamic->stencil_reference.front,
- dynamic->stencil_reference.back);
- }
-
- if (cmd->state.dirty &
- (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_VERTEX_BUFFERS)) {
- for (uint32_t i = 0; i < pipeline->vi.count; i++) {
- const uint32_t binding = pipeline->vi.bindings[i];
- const uint32_t stride = pipeline->vi.strides[i];
- const struct tu_buffer *buf = cmd->state.vb.buffers[binding];
- const VkDeviceSize offset = buf->bo_offset +
- cmd->state.vb.offsets[binding] +
- pipeline->vi.offsets[i];
- const VkDeviceSize size =
- offset < buf->bo->size ? buf->bo->size - offset : 0;
-
- tu_cs_emit_regs(cs,
- A6XX_VFD_FETCH_BASE(i, .bo = buf->bo, .bo_offset = offset),
- A6XX_VFD_FETCH_SIZE(i, size),
- A6XX_VFD_FETCH_STRIDE(i, stride));
- }
- }
+ tu_emit_cache_flush_renderpass(cmd, cs);
- if (cmd->state.dirty & TU_CMD_DIRTY_PIPELINE) {
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_PROGRAM,
- .enable_mask = ENABLE_DRAW,
- .ib = pipeline->program.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_PROGRAM_BINNING,
- .enable_mask = CP_SET_DRAW_STATE__0_BINNING,
- .ib = pipeline->program.binning_state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VI,
- .enable_mask = ENABLE_DRAW,
- .ib = pipeline->vi.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VI_BINNING,
- .enable_mask = CP_SET_DRAW_STATE__0_BINNING,
- .ib = pipeline->vi.binning_state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VP,
- .enable_mask = ENABLE_ALL,
- .ib = pipeline->vp.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_RAST,
- .enable_mask = ENABLE_ALL,
- .ib = pipeline->rast.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_DS,
- .enable_mask = ENABLE_ALL,
- .ib = pipeline->ds.state_ib,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_BLEND,
- .enable_mask = ENABLE_ALL,
- .ib = pipeline->blend.state_ib,
- };
- }
-
- if (cmd->state.dirty &
- (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DESCRIPTOR_SETS | TU_CMD_DIRTY_PUSH_CONSTANTS)) {
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VS_CONST,
- .enable_mask = ENABLE_ALL,
- .ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_VERTEX)
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_FS_CONST,
- .enable_mask = ENABLE_DRAW,
- .ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_FRAGMENT)
- };
- }
-
- if (cmd->state.dirty & TU_CMD_DIRTY_STREAMOUT_BUFFERS)
- tu6_emit_streamout(cmd, cs);
-
- if (cmd->state.dirty &
- (TU_CMD_DIRTY_PIPELINE | TU_CMD_DIRTY_DESCRIPTOR_SETS)) {
- struct tu_cs_entry vs_tex, fs_tex_sysmem, fs_tex_gmem, fs_ibo;
+ /* TODO lrz */
- result = tu6_emit_textures(cmd, pipeline, descriptors_state,
- MESA_SHADER_VERTEX, &vs_tex, false);
+ tu_cs_emit_regs(cs, A6XX_PC_PRIMITIVE_CNTL_0(
+ .primitive_restart =
+ pipeline->ia.primitive_restart && indexed,
+ .tess_upper_left_domain_origin =
+ pipeline->tess.upper_left_domain_origin));
+
+ if (cmd->state.dirty & TU_CMD_DIRTY_SHADER_CONSTS) {
+ cmd->state.shader_const[MESA_SHADER_VERTEX] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_VERTEX);
+ cmd->state.shader_const[MESA_SHADER_TESS_CTRL] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_TESS_CTRL);
+ cmd->state.shader_const[MESA_SHADER_TESS_EVAL] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_TESS_EVAL);
+ cmd->state.shader_const[MESA_SHADER_GEOMETRY] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_GEOMETRY);
+ cmd->state.shader_const[MESA_SHADER_FRAGMENT] =
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_FRAGMENT);
+ }
+
+ if (cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS)
+ cmd->state.vertex_buffers = tu6_emit_vertex_buffers(cmd, pipeline);
+
+ bool has_tess =
+ pipeline->active_stages & VK_SHADER_STAGE_TESSELLATION_CONTROL_BIT;
+ struct tu_draw_state tess_consts = {};
+ if (has_tess) {
+ uint64_t tess_factor_iova = 0;
+
+ cmd->has_tess = true;
+ result = tu6_emit_tess_consts(cmd, draw_count, pipeline, &tess_consts, &tess_factor_iova);
if (result != VK_SUCCESS)
return result;
- /* TODO: we could emit just one texture descriptor draw state when there
- * are no input attachments, which is the most common case. We could
- * also split out the sampler state, which doesn't change even for input
- * attachments.
+ /* this sequence matches what the blob does before every tess draw
+ * PC_TESSFACTOR_ADDR_LO is a non-context register and needs a wfi
+ * before writing to it
*/
- result = tu6_emit_textures(cmd, pipeline, descriptors_state,
- MESA_SHADER_FRAGMENT, &fs_tex_sysmem, true);
- if (result != VK_SUCCESS)
- return result;
-
- result = tu6_emit_textures(cmd, pipeline, descriptors_state,
- MESA_SHADER_FRAGMENT, &fs_tex_gmem, false);
- if (result != VK_SUCCESS)
- return result;
+ tu_cs_emit_wfi(cs);
- result = tu6_emit_ibo(cmd, pipeline, descriptors_state,
- MESA_SHADER_FRAGMENT, &fs_ibo);
- if (result != VK_SUCCESS)
- return result;
+ tu_cs_emit_pkt4(cs, REG_A6XX_PC_TESSFACTOR_ADDR_LO, 2);
+ tu_cs_emit_qw(cs, tess_factor_iova);
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VS_TEX,
- .enable_mask = ENABLE_ALL,
- .ib = vs_tex,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_FS_TEX_GMEM,
- .enable_mask = CP_SET_DRAW_STATE__0_GMEM,
- .ib = fs_tex_gmem,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_FS_TEX_SYSMEM,
- .enable_mask = CP_SET_DRAW_STATE__0_SYSMEM,
- .ib = fs_tex_sysmem,
- };
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_FS_IBO,
- .enable_mask = ENABLE_DRAW,
- .ib = fs_ibo,
- };
+ tu_cs_emit_pkt7(cs, CP_SET_SUBDRAW_SIZE, 1);
+ tu_cs_emit(cs, draw_count);
}
- struct tu_cs_entry vs_params;
- result = tu6_emit_vs_params(cmd, draw, &vs_params);
- if (result != VK_SUCCESS)
- return result;
-
- draw_state_groups[draw_state_group_count++] =
- (struct tu_draw_state_group) {
- .id = TU_DRAW_STATE_VS_PARAMS,
- .enable_mask = ENABLE_ALL,
- .ib = vs_params,
- };
-
- tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * draw_state_group_count);
- for (uint32_t i = 0; i < draw_state_group_count; i++) {
- const struct tu_draw_state_group *group = &draw_state_groups[i];
- debug_assert((group->enable_mask & ~ENABLE_ALL) == 0);
- uint32_t cp_set_draw_state =
- CP_SET_DRAW_STATE__0_COUNT(group->ib.size / 4) |
- group->enable_mask |
- CP_SET_DRAW_STATE__0_GROUP_ID(group->id);
- uint64_t iova;
- if (group->ib.size) {
- iova = group->ib.bo->iova + group->ib.offset;
- } else {
- cp_set_draw_state |= CP_SET_DRAW_STATE__0_DISABLE;
- iova = 0;
+ /* for the first draw in a renderpass, re-emit all the draw states
+ *
+ * and if a draw-state disabling path (CmdClearAttachments 3D fallback) was
+ * used, then draw states must be re-emitted. note however this only happens
+ * in the sysmem path, so this can be skipped this for the gmem path (TODO)
+ *
+ * the two input attachment states are excluded because secondary command
+ * buffer doesn't have a state ib to restore it, and not re-emitting them
+ * is OK since CmdClearAttachments won't disable/overwrite them
+ */
+ if (cmd->state.dirty & TU_CMD_DIRTY_DRAW_STATE) {
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * (TU_DRAW_STATE_COUNT - 2));
+
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM, pipeline->program.state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_PROGRAM_BINNING, pipeline->program.binning_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_TESS, tess_consts);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VI, pipeline->vi.state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VI_BINNING, pipeline->vi.binning_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_RAST, pipeline->rast_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DS, pipeline->ds_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_BLEND, pipeline->blend_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VS_CONST, cmd->state.shader_const[MESA_SHADER_VERTEX]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_HS_CONST, cmd->state.shader_const[MESA_SHADER_TESS_CTRL]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DS_CONST, cmd->state.shader_const[MESA_SHADER_TESS_EVAL]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_GS_CONST, cmd->state.shader_const[MESA_SHADER_GEOMETRY]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_FS_CONST, cmd->state.shader_const[MESA_SHADER_FRAGMENT]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DESC_SETS, cmd->state.desc_sets);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DESC_SETS_LOAD, pipeline->load_state);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VB, cmd->state.vertex_buffers);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VS_PARAMS, cmd->state.vs_params);
+
+ for (uint32_t i = 0; i < ARRAY_SIZE(cmd->state.dynamic_state); i++) {
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DYNAMIC + i,
+ ((pipeline->dynamic_state_mask & BIT(i)) ?
+ cmd->state.dynamic_state[i] :
+ pipeline->dynamic_state[i]));
}
+ } else {
- tu_cs_emit(cs, cp_set_draw_state);
- tu_cs_emit_qw(cs, iova);
+ /* emit draw states that were just updated
+ * note we eventually don't want to have to emit anything here
+ */
+ uint32_t draw_state_count =
+ has_tess +
+ ((cmd->state.dirty & TU_CMD_DIRTY_SHADER_CONSTS) ? 5 : 0) +
+ ((cmd->state.dirty & TU_CMD_DIRTY_DESC_SETS_LOAD) ? 1 : 0) +
+ ((cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS) ? 1 : 0) +
+ 1; /* vs_params */
+
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3 * draw_state_count);
+
+ /* We may need to re-emit tess consts if the current draw call is
+ * sufficiently larger than the last draw call. */
+ if (has_tess)
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_TESS, tess_consts);
+ if (cmd->state.dirty & TU_CMD_DIRTY_SHADER_CONSTS) {
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VS_CONST, cmd->state.shader_const[MESA_SHADER_VERTEX]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_HS_CONST, cmd->state.shader_const[MESA_SHADER_TESS_CTRL]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DS_CONST, cmd->state.shader_const[MESA_SHADER_TESS_EVAL]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_GS_CONST, cmd->state.shader_const[MESA_SHADER_GEOMETRY]);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_FS_CONST, cmd->state.shader_const[MESA_SHADER_FRAGMENT]);
+ }
+ if (cmd->state.dirty & TU_CMD_DIRTY_DESC_SETS_LOAD)
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_DESC_SETS_LOAD, pipeline->load_state);
+ if (cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS)
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VB, cmd->state.vertex_buffers);
+ tu_cs_emit_draw_state(cs, TU_DRAW_STATE_VS_PARAMS, cmd->state.vs_params);
}
tu_cs_sanity_check(cs);
- /* track BOs */
- if (cmd->state.dirty & TU_CMD_DIRTY_VERTEX_BUFFERS) {
- for (uint32_t i = 0; i < MAX_VBS; i++) {
- const struct tu_buffer *buf = cmd->state.vb.buffers[i];
- if (buf)
- tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
- }
- }
- if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS) {
- unsigned i;
- for_each_bit(i, descriptors_state->valid) {
- struct tu_descriptor_set *set = descriptors_state->sets[i];
- for (unsigned j = 0; j < set->layout->buffer_count; ++j)
- if (set->descriptors[j]) {
- tu_bo_list_add(&cmd->bo_list, set->descriptors[j],
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- }
- }
- }
- if (cmd->state.dirty & TU_CMD_DIRTY_STREAMOUT_BUFFERS) {
- for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
- const struct tu_buffer *buf = cmd->state.streamout_buf.buffers[i];
- if (buf) {
- tu_bo_list_add(&cmd->bo_list, buf->bo,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- }
- }
- }
-
- /* Fragment shader state overwrites compute shader state, so flag the
- * compute pipeline for re-emit.
+ /* There are too many graphics dirty bits to list here, so just list the
+ * bits to preserve instead. The only things not emitted here are
+ * compute-related state.
*/
- cmd->state.dirty = TU_CMD_DIRTY_COMPUTE_PIPELINE;
+ cmd->state.dirty &= TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD;
return VK_SUCCESS;
}
-static void
-tu6_emit_draw_indirect(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_draw_info *draw)
+static uint32_t
+tu_draw_initiator(struct tu_cmd_buffer *cmd, enum pc_di_src_sel src_sel)
{
- const enum pc_di_primtype primtype = cmd->state.pipeline->ia.primtype;
- bool has_gs = cmd->state.pipeline->active_stages &
- VK_SHADER_STAGE_GEOMETRY_BIT;
-
- tu_cs_emit_regs(cs,
- A6XX_VFD_INDEX_OFFSET(draw->vertex_offset),
- A6XX_VFD_INSTANCE_START_OFFSET(draw->first_instance));
-
- if (draw->indexed) {
- const enum a4xx_index_size index_size =
- tu6_index_size(cmd->state.index_type);
- const uint32_t index_bytes =
- (cmd->state.index_type == VK_INDEX_TYPE_UINT32) ? 4 : 2;
- const struct tu_buffer *index_buf = cmd->state.index_buffer;
- unsigned max_indicies =
- (index_buf->size - cmd->state.index_offset) / index_bytes;
-
- const uint32_t cp_draw_indx =
- CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
- CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_DMA) |
- CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(index_size) |
- CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) |
- COND(has_gs, CP_DRAW_INDX_OFFSET_0_GS_ENABLE) | 0x2000;
-
- tu_cs_emit_pkt7(cs, CP_DRAW_INDX_INDIRECT, 6);
- tu_cs_emit(cs, cp_draw_indx);
- tu_cs_emit_qw(cs, index_buf->bo->iova + cmd->state.index_offset);
- tu_cs_emit(cs, A5XX_CP_DRAW_INDX_INDIRECT_3_MAX_INDICES(max_indicies));
- tu_cs_emit_qw(cs, draw->indirect->bo->iova + draw->indirect_offset);
- } else {
- const uint32_t cp_draw_indx =
- CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
- CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_AUTO_INDEX) |
- CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) |
- COND(has_gs, CP_DRAW_INDX_OFFSET_0_GS_ENABLE) | 0x2000;
-
- tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT, 3);
- tu_cs_emit(cs, cp_draw_indx);
- tu_cs_emit_qw(cs, draw->indirect->bo->iova + draw->indirect_offset);
+ const struct tu_pipeline *pipeline = cmd->state.pipeline;
+ uint32_t initiator =
+ CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(pipeline->ia.primtype) |
+ CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(src_sel) |
+ CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(cmd->state.index_size) |
+ CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY);
+
+ if (pipeline->active_stages & VK_SHADER_STAGE_GEOMETRY_BIT)
+ initiator |= CP_DRAW_INDX_OFFSET_0_GS_ENABLE;
+
+ switch (pipeline->tess.patch_type) {
+ case IR3_TESS_TRIANGLES:
+ initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_TRIANGLES) |
+ CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
+ break;
+ case IR3_TESS_ISOLINES:
+ initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_ISOLINES) |
+ CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
+ break;
+ case IR3_TESS_NONE:
+ initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_QUADS);
+ break;
+ case IR3_TESS_QUADS:
+ initiator |= CP_DRAW_INDX_OFFSET_0_PATCH_TYPE(TESS_QUADS) |
+ CP_DRAW_INDX_OFFSET_0_TESS_ENABLE;
+ break;
}
-
- tu_bo_list_add(&cmd->bo_list, draw->indirect->bo, MSM_SUBMIT_BO_READ);
+ return initiator;
}
-static void
-tu6_emit_draw_direct(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_draw_info *draw)
+
+static uint32_t
+vs_params_offset(struct tu_cmd_buffer *cmd)
{
+ const struct tu_program_descriptor_linkage *link =
+ &cmd->state.pipeline->program.link[MESA_SHADER_VERTEX];
+ const struct ir3_const_state *const_state = &link->const_state;
- const enum pc_di_primtype primtype = cmd->state.pipeline->ia.primtype;
- bool has_gs = cmd->state.pipeline->active_stages &
- VK_SHADER_STAGE_GEOMETRY_BIT;
+ if (const_state->offsets.driver_param >= link->constlen)
+ return 0;
- tu_cs_emit_regs(cs,
- A6XX_VFD_INDEX_OFFSET(draw->vertex_offset),
- A6XX_VFD_INSTANCE_START_OFFSET(draw->first_instance));
-
- /* TODO hw binning */
- if (draw->indexed) {
- const enum a4xx_index_size index_size =
- tu6_index_size(cmd->state.index_type);
- const uint32_t index_bytes =
- (cmd->state.index_type == VK_INDEX_TYPE_UINT32) ? 4 : 2;
- const struct tu_buffer *buf = cmd->state.index_buffer;
- const VkDeviceSize offset = buf->bo_offset + cmd->state.index_offset +
- index_bytes * draw->first_index;
- const uint32_t size = index_bytes * draw->count;
-
- const uint32_t cp_draw_indx =
- CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
- CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_DMA) |
- CP_DRAW_INDX_OFFSET_0_INDEX_SIZE(index_size) |
- CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) |
- COND(has_gs, CP_DRAW_INDX_OFFSET_0_GS_ENABLE) | 0x2000;
-
- tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
- tu_cs_emit(cs, cp_draw_indx);
- tu_cs_emit(cs, draw->instance_count);
- tu_cs_emit(cs, draw->count);
- tu_cs_emit(cs, 0x0); /* XXX */
- tu_cs_emit_qw(cs, buf->bo->iova + offset);
- tu_cs_emit(cs, size);
- } else {
- const uint32_t cp_draw_indx =
- CP_DRAW_INDX_OFFSET_0_PRIM_TYPE(primtype) |
- CP_DRAW_INDX_OFFSET_0_SOURCE_SELECT(DI_SRC_SEL_AUTO_INDEX) |
- CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) |
- COND(has_gs, CP_DRAW_INDX_OFFSET_0_GS_ENABLE) | 0x2000;
-
- tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
- tu_cs_emit(cs, cp_draw_indx);
- tu_cs_emit(cs, draw->instance_count);
- tu_cs_emit(cs, draw->count);
- }
+ /* this layout is required by CP_DRAW_INDIRECT_MULTI */
+ STATIC_ASSERT(IR3_DP_DRAWID == 0);
+ STATIC_ASSERT(IR3_DP_VTXID_BASE == 1);
+ STATIC_ASSERT(IR3_DP_INSTID_BASE == 2);
+
+ /* 0 means disabled for CP_DRAW_INDIRECT_MULTI */
+ assert(const_state->offsets.driver_param != 0);
+
+ return const_state->offsets.driver_param;
}
-static void
-tu_draw(struct tu_cmd_buffer *cmd, const struct tu_draw_info *draw)
+static struct tu_draw_state
+tu6_emit_vs_params(struct tu_cmd_buffer *cmd,
+ uint32_t vertex_offset,
+ uint32_t first_instance)
{
- struct tu_cs *cs = &cmd->draw_cs;
- VkResult result;
+ uint32_t offset = vs_params_offset(cmd);
- result = tu6_bind_draw_states(cmd, cs, draw);
+ struct tu_cs cs;
+ VkResult result = tu_cs_begin_sub_stream(&cmd->sub_cs, 3 + (offset ? 8 : 0), &cs);
if (result != VK_SUCCESS) {
cmd->record_result = result;
- return;
+ return (struct tu_draw_state) {};
}
- if (draw->indirect)
- tu6_emit_draw_indirect(cmd, cs, draw);
- else
- tu6_emit_draw_direct(cmd, cs, draw);
+ /* TODO: don't make a new draw state when it doesn't change */
- if (cmd->state.streamout_enabled) {
- for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
- if (cmd->state.streamout_enabled & (1 << i))
- tu6_emit_event_write(cmd, cs, FLUSH_SO_0 + i, false);
- }
- }
+ tu_cs_emit_regs(&cs,
+ A6XX_VFD_INDEX_OFFSET(vertex_offset),
+ A6XX_VFD_INSTANCE_START_OFFSET(first_instance));
- cmd->wait_for_idle = true;
+ if (offset) {
+ tu_cs_emit_pkt7(&cs, CP_LOAD_STATE6_GEOM, 3 + 4);
+ tu_cs_emit(&cs, CP_LOAD_STATE6_0_DST_OFF(offset) |
+ CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS) |
+ CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
+ CP_LOAD_STATE6_0_STATE_BLOCK(SB6_VS_SHADER) |
+ CP_LOAD_STATE6_0_NUM_UNIT(1));
+ tu_cs_emit(&cs, 0);
+ tu_cs_emit(&cs, 0);
- tu_cs_sanity_check(cs);
+ tu_cs_emit(&cs, 0);
+ tu_cs_emit(&cs, vertex_offset);
+ tu_cs_emit(&cs, first_instance);
+ tu_cs_emit(&cs, 0);
+ }
+
+ struct tu_cs_entry entry = tu_cs_end_sub_stream(&cmd->sub_cs, &cs);
+ return (struct tu_draw_state) {entry.bo->iova + entry.offset, entry.size / 4};
}
void
uint32_t firstVertex,
uint32_t firstInstance)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- struct tu_draw_info info = {};
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- info.count = vertexCount;
- info.instance_count = instanceCount;
- info.first_instance = firstInstance;
- info.vertex_offset = firstVertex;
+ cmd->state.vs_params = tu6_emit_vs_params(cmd, firstVertex, firstInstance);
- tu_draw(cmd_buffer, &info);
+ tu6_draw_common(cmd, cs, false, vertexCount);
+
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
+ tu_cs_emit(cs, instanceCount);
+ tu_cs_emit(cs, vertexCount);
}
void
int32_t vertexOffset,
uint32_t firstInstance)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- struct tu_draw_info info = {};
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- info.indexed = true;
- info.count = indexCount;
- info.instance_count = instanceCount;
- info.first_index = firstIndex;
- info.vertex_offset = vertexOffset;
- info.first_instance = firstInstance;
+ cmd->state.vs_params = tu6_emit_vs_params(cmd, vertexOffset, firstInstance);
- tu_draw(cmd_buffer, &info);
+ tu6_draw_common(cmd, cs, true, indexCount);
+
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
+ tu_cs_emit(cs, instanceCount);
+ tu_cs_emit(cs, indexCount);
+ tu_cs_emit(cs, firstIndex);
+ tu_cs_emit_qw(cs, cmd->state.index_va);
+ tu_cs_emit(cs, cmd->state.max_index_count);
+}
+
+/* Various firmware bugs/inconsistencies mean that some indirect draw opcodes
+ * do not wait for WFI's to complete before executing. Add a WAIT_FOR_ME if
+ * pending for these opcodes. This may result in a few extra WAIT_FOR_ME's
+ * with these opcodes, but the alternative would add unnecessary WAIT_FOR_ME's
+ * before draw opcodes that don't need it.
+ */
+static void
+draw_wfm(struct tu_cmd_buffer *cmd)
+{
+ cmd->state.renderpass_cache.flush_bits |=
+ cmd->state.renderpass_cache.pending_flush_bits & TU_CMD_FLAG_WAIT_FOR_ME;
+ cmd->state.renderpass_cache.pending_flush_bits &= ~TU_CMD_FLAG_WAIT_FOR_ME;
}
void
uint32_t drawCount,
uint32_t stride)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
- struct tu_draw_info info = {};
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ TU_FROM_HANDLE(tu_buffer, buf, _buffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- info.count = drawCount;
- info.indirect = buffer;
- info.indirect_offset = offset;
- info.stride = stride;
+ cmd->state.vs_params = (struct tu_draw_state) {};
+
+ /* The latest known a630_sqe.fw fails to wait for WFI before reading the
+ * indirect buffer when using CP_DRAW_INDIRECT_MULTI, so we have to fall
+ * back to CP_WAIT_FOR_ME except for a650 which has a fixed firmware.
+ *
+ * TODO: There may be newer a630_sqe.fw released in the future which fixes
+ * this, if so we should detect it and avoid this workaround.
+ */
+ if (cmd->device->physical_device->gpu_id != 650)
+ draw_wfm(cmd);
+
+ tu6_draw_common(cmd, cs, false, 0);
- tu_draw(cmd_buffer, &info);
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT_MULTI, 6);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
+ tu_cs_emit(cs, A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_NORMAL) |
+ A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(vs_params_offset(cmd)));
+ tu_cs_emit(cs, drawCount);
+ tu_cs_emit_qw(cs, buf->bo->iova + buf->bo_offset + offset);
+ tu_cs_emit(cs, stride);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
}
void
uint32_t drawCount,
uint32_t stride)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- TU_FROM_HANDLE(tu_buffer, buffer, _buffer);
- struct tu_draw_info info = {};
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ TU_FROM_HANDLE(tu_buffer, buf, _buffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- info.indexed = true;
- info.count = drawCount;
- info.indirect = buffer;
- info.indirect_offset = offset;
- info.stride = stride;
+ cmd->state.vs_params = (struct tu_draw_state) {};
+
+ if (cmd->device->physical_device->gpu_id != 650)
+ draw_wfm(cmd);
- tu_draw(cmd_buffer, &info);
+ tu6_draw_common(cmd, cs, true, 0);
+
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT_MULTI, 9);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
+ tu_cs_emit(cs, A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_INDEXED) |
+ A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(vs_params_offset(cmd)));
+ tu_cs_emit(cs, drawCount);
+ tu_cs_emit_qw(cs, cmd->state.index_va);
+ tu_cs_emit(cs, cmd->state.max_index_count);
+ tu_cs_emit_qw(cs, buf->bo->iova + buf->bo_offset + offset);
+ tu_cs_emit(cs, stride);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+}
+
+void
+tu_CmdDrawIndirectCount(VkCommandBuffer commandBuffer,
+ VkBuffer _buffer,
+ VkDeviceSize offset,
+ VkBuffer countBuffer,
+ VkDeviceSize countBufferOffset,
+ uint32_t drawCount,
+ uint32_t stride)
+{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ TU_FROM_HANDLE(tu_buffer, buf, _buffer);
+ TU_FROM_HANDLE(tu_buffer, count_buf, countBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ cmd->state.vs_params = (struct tu_draw_state) {};
+
+ /* It turns out that the firmware we have for a650 only partially fixed the
+ * problem with CP_DRAW_INDIRECT_MULTI not waiting for WFI's to complete
+ * before reading indirect parameters. It waits for WFI's before reading
+ * the draw parameters, but after reading the indirect count :(.
+ */
+ draw_wfm(cmd);
+
+ tu6_draw_common(cmd, cs, false, 0);
+
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT_MULTI, 8);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_INDEX));
+ tu_cs_emit(cs, A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_INDIRECT_COUNT) |
+ A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(vs_params_offset(cmd)));
+ tu_cs_emit(cs, drawCount);
+ tu_cs_emit_qw(cs, buf->bo->iova + buf->bo_offset + offset);
+ tu_cs_emit_qw(cs, count_buf->bo->iova + count_buf->bo_offset + countBufferOffset);
+ tu_cs_emit(cs, stride);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+ tu_bo_list_add(&cmd->bo_list, count_buf->bo, MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+}
+
+void
+tu_CmdDrawIndexedIndirectCount(VkCommandBuffer commandBuffer,
+ VkBuffer _buffer,
+ VkDeviceSize offset,
+ VkBuffer countBuffer,
+ VkDeviceSize countBufferOffset,
+ uint32_t drawCount,
+ uint32_t stride)
+{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ TU_FROM_HANDLE(tu_buffer, buf, _buffer);
+ TU_FROM_HANDLE(tu_buffer, count_buf, countBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ cmd->state.vs_params = (struct tu_draw_state) {};
+
+ draw_wfm(cmd);
+
+ tu6_draw_common(cmd, cs, true, 0);
+
+ tu_cs_emit_pkt7(cs, CP_DRAW_INDIRECT_MULTI, 11);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_DMA));
+ tu_cs_emit(cs, A6XX_CP_DRAW_INDIRECT_MULTI_1_OPCODE(INDIRECT_OP_INDIRECT_COUNT_INDEXED) |
+ A6XX_CP_DRAW_INDIRECT_MULTI_1_DST_OFF(vs_params_offset(cmd)));
+ tu_cs_emit(cs, drawCount);
+ tu_cs_emit_qw(cs, cmd->state.index_va);
+ tu_cs_emit(cs, cmd->state.max_index_count);
+ tu_cs_emit_qw(cs, buf->bo->iova + buf->bo_offset + offset);
+ tu_cs_emit_qw(cs, count_buf->bo->iova + count_buf->bo_offset + countBufferOffset);
+ tu_cs_emit(cs, stride);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
+ tu_bo_list_add(&cmd->bo_list, count_buf->bo, MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
}
void tu_CmdDrawIndirectByteCountEXT(VkCommandBuffer commandBuffer,
uint32_t counterOffset,
uint32_t vertexStride)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
- TU_FROM_HANDLE(tu_buffer, buffer, _counterBuffer);
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ TU_FROM_HANDLE(tu_buffer, buf, _counterBuffer);
+ struct tu_cs *cs = &cmd->draw_cs;
- struct tu_draw_info info = {};
+ /* All known firmware versions do not wait for WFI's with CP_DRAW_AUTO.
+ * Plus, for the common case where the counter buffer is written by
+ * vkCmdEndTransformFeedback, we need to wait for the CP_WAIT_MEM_WRITES to
+ * complete which means we need a WAIT_FOR_ME anyway.
+ */
+ draw_wfm(cmd);
+
+ cmd->state.vs_params = tu6_emit_vs_params(cmd, 0, firstInstance);
- info.instance_count = instanceCount;
- info.first_instance = firstInstance;
- info.streamout_buffer = buffer;
- info.streamout_buffer_offset = counterBufferOffset;
- info.stride = vertexStride;
+ tu6_draw_common(cmd, cs, false, 0);
- tu_draw(cmd_buffer, &info);
+ tu_cs_emit_pkt7(cs, CP_DRAW_AUTO, 6);
+ tu_cs_emit(cs, tu_draw_initiator(cmd, DI_SRC_SEL_AUTO_XFB));
+ tu_cs_emit(cs, instanceCount);
+ tu_cs_emit_qw(cs, buf->bo->iova + buf->bo_offset + counterBufferOffset);
+ tu_cs_emit(cs, counterOffset);
+ tu_cs_emit(cs, vertexStride);
+
+ tu_bo_list_add(&cmd->bo_list, buf->bo, MSM_SUBMIT_BO_READ);
}
struct tu_dispatch_info
struct tu_pipeline *pipeline = cmd->state.compute_pipeline;
struct tu_descriptor_state *descriptors_state =
&cmd->descriptors[VK_PIPELINE_BIND_POINT_COMPUTE];
- VkResult result;
- if (cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_PIPELINE)
- tu_cs_emit_ib(cs, &pipeline->program.state_ib);
-
- struct tu_cs_entry ib;
+ /* TODO: We could probably flush less if we add a compute_flush_bits
+ * bitfield.
+ */
+ tu_emit_cache_flush(cmd, cs);
- ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_COMPUTE);
- if (ib.size)
- tu_cs_emit_ib(cs, &ib);
+ /* note: no reason to have this in a separate IB */
+ tu_cs_emit_state_ib(cs,
+ tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_COMPUTE));
tu_emit_compute_driver_params(cs, pipeline, info);
- result = tu6_emit_textures(cmd, pipeline, descriptors_state,
- MESA_SHADER_COMPUTE, &ib, false);
- if (result != VK_SUCCESS) {
- cmd->record_result = result;
- return;
- }
-
- if (ib.size)
- tu_cs_emit_ib(cs, &ib);
-
- result = tu6_emit_ibo(cmd, pipeline, descriptors_state, MESA_SHADER_COMPUTE, &ib);
- if (result != VK_SUCCESS) {
- cmd->record_result = result;
- return;
- }
-
- if (ib.size)
- tu_cs_emit_ib(cs, &ib);
-
- /* track BOs */
- if (cmd->state.dirty & TU_CMD_DIRTY_DESCRIPTOR_SETS) {
- unsigned i;
- for_each_bit(i, descriptors_state->valid) {
- struct tu_descriptor_set *set = descriptors_state->sets[i];
- for (unsigned j = 0; j < set->layout->buffer_count; ++j)
- if (set->descriptors[j]) {
- tu_bo_list_add(&cmd->bo_list, set->descriptors[j],
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- }
- }
- }
+ if (cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD)
+ tu_cs_emit_state_ib(cs, pipeline->load_state);
- /* Compute shader state overwrites fragment shader state, so we flag the
- * graphics pipeline for re-emit.
- */
- cmd->state.dirty = TU_CMD_DIRTY_PIPELINE;
+ cmd->state.dirty &= ~TU_CMD_DIRTY_COMPUTE_DESC_SETS_LOAD;
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_COMPUTE));
}
tu_cs_emit_wfi(cs);
-
- tu6_emit_cache_flush(cmd, cs);
}
void
tu_cs_discard_entries(&cmd_buffer->draw_epilogue_cs);
tu_cs_begin(&cmd_buffer->draw_epilogue_cs);
+ cmd_buffer->state.cache.pending_flush_bits |=
+ cmd_buffer->state.renderpass_cache.pending_flush_bits;
+ tu_subpass_barrier(cmd_buffer, &cmd_buffer->state.pass->end_barrier, true);
+
cmd_buffer->state.pass = NULL;
cmd_buffer->state.subpass = NULL;
cmd_buffer->state.framebuffer = NULL;
};
static void
-tu_barrier(struct tu_cmd_buffer *cmd_buffer,
+tu_barrier(struct tu_cmd_buffer *cmd,
uint32_t memoryBarrierCount,
const VkMemoryBarrier *pMemoryBarriers,
uint32_t bufferMemoryBarrierCount,
const VkImageMemoryBarrier *pImageMemoryBarriers,
const struct tu_barrier_info *info)
{
+ struct tu_cs *cs = cmd->state.pass ? &cmd->draw_cs : &cmd->cs;
+ VkAccessFlags srcAccessMask = 0;
+ VkAccessFlags dstAccessMask = 0;
+
+ for (uint32_t i = 0; i < memoryBarrierCount; i++) {
+ srcAccessMask |= pMemoryBarriers[i].srcAccessMask;
+ dstAccessMask |= pMemoryBarriers[i].dstAccessMask;
+ }
+
+ for (uint32_t i = 0; i < bufferMemoryBarrierCount; i++) {
+ srcAccessMask |= pBufferMemoryBarriers[i].srcAccessMask;
+ dstAccessMask |= pBufferMemoryBarriers[i].dstAccessMask;
+ }
+
+ enum tu_cmd_access_mask src_flags = 0;
+ enum tu_cmd_access_mask dst_flags = 0;
+
+ for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
+ TU_FROM_HANDLE(tu_image, image, pImageMemoryBarriers[i].image);
+ VkImageLayout old_layout = pImageMemoryBarriers[i].oldLayout;
+ /* For non-linear images, PREINITIALIZED is the same as UNDEFINED */
+ if (old_layout == VK_IMAGE_LAYOUT_UNDEFINED ||
+ (image->tiling != VK_IMAGE_TILING_LINEAR &&
+ old_layout == VK_IMAGE_LAYOUT_PREINITIALIZED)) {
+ /* The underlying memory for this image may have been used earlier
+ * within the same queue submission for a different image, which
+ * means that there may be old, stale cache entries which are in the
+ * "wrong" location, which could cause problems later after writing
+ * to the image. We don't want these entries being flushed later and
+ * overwriting the actual image, so we need to flush the CCU.
+ */
+ src_flags |= TU_ACCESS_CCU_COLOR_INCOHERENT_WRITE;
+ }
+ srcAccessMask |= pImageMemoryBarriers[i].srcAccessMask;
+ dstAccessMask |= pImageMemoryBarriers[i].dstAccessMask;
+ }
+
+ /* Inside a renderpass, we don't know yet whether we'll be using sysmem
+ * so we have to use the sysmem flushes.
+ */
+ bool gmem = cmd->state.ccu_state == TU_CMD_CCU_GMEM &&
+ !cmd->state.pass;
+ src_flags |= vk2tu_access(srcAccessMask, gmem);
+ dst_flags |= vk2tu_access(dstAccessMask, gmem);
+
+ struct tu_cache_state *cache =
+ cmd->state.pass ? &cmd->state.renderpass_cache : &cmd->state.cache;
+ tu_flush_for_access(cache, src_flags, dst_flags);
+
+ for (uint32_t i = 0; i < info->eventCount; i++) {
+ TU_FROM_HANDLE(tu_event, event, info->pEvents[i]);
+
+ tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_READ);
+
+ tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
+ tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
+ CP_WAIT_REG_MEM_0_POLL_MEMORY);
+ tu_cs_emit_qw(cs, event->bo.iova); /* POLL_ADDR_LO/HI */
+ tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(1));
+ tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0u));
+ tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(20));
+ }
}
void
tu_CmdPipelineBarrier(VkCommandBuffer commandBuffer,
VkPipelineStageFlags srcStageMask,
- VkPipelineStageFlags destStageMask,
- VkBool32 byRegion,
+ VkPipelineStageFlags dstStageMask,
+ VkDependencyFlags dependencyFlags,
uint32_t memoryBarrierCount,
const VkMemoryBarrier *pMemoryBarriers,
uint32_t bufferMemoryBarrierCount,
}
static void
-write_event(struct tu_cmd_buffer *cmd, struct tu_event *event, unsigned value)
+write_event(struct tu_cmd_buffer *cmd, struct tu_event *event,
+ VkPipelineStageFlags stageMask, unsigned value)
{
struct tu_cs *cs = &cmd->cs;
- tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_WRITE);
+ /* vkCmdSetEvent/vkCmdResetEvent cannot be called inside a render pass */
+ assert(!cmd->state.pass);
- /* TODO: any flush required before/after ? */
+ tu_emit_cache_flush(cmd, cs);
- tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
- tu_cs_emit_qw(cs, event->bo.iova); /* ADDR_LO/HI */
- tu_cs_emit(cs, value);
+ tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_WRITE);
+
+ /* Flags that only require a top-of-pipe event. DrawIndirect parameters are
+ * read by the CP, so the draw indirect stage counts as top-of-pipe too.
+ */
+ VkPipelineStageFlags top_of_pipe_flags =
+ VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT |
+ VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT;
+
+ if (!(stageMask & ~top_of_pipe_flags)) {
+ tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
+ tu_cs_emit_qw(cs, event->bo.iova); /* ADDR_LO/HI */
+ tu_cs_emit(cs, value);
+ } else {
+ /* Use a RB_DONE_TS event to wait for everything to complete. */
+ tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 4);
+ tu_cs_emit(cs, CP_EVENT_WRITE_0_EVENT(RB_DONE_TS));
+ tu_cs_emit_qw(cs, event->bo.iova);
+ tu_cs_emit(cs, value);
+ }
}
void
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_event, event, _event);
- write_event(cmd, event, 1);
+ write_event(cmd, event, stageMask, 1);
}
void
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_event, event, _event);
- write_event(cmd, event, 0);
+ write_event(cmd, event, stageMask, 0);
}
void
const VkImageMemoryBarrier *pImageMemoryBarriers)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
- struct tu_cs *cs = &cmd->cs;
-
- /* TODO: any flush required before/after? (CP_WAIT_FOR_ME?) */
-
- for (uint32_t i = 0; i < eventCount; i++) {
- TU_FROM_HANDLE(tu_event, event, pEvents[i]);
+ struct tu_barrier_info info;
- tu_bo_list_add(&cmd->bo_list, &event->bo, MSM_SUBMIT_BO_READ);
+ info.eventCount = eventCount;
+ info.pEvents = pEvents;
+ info.srcStageMask = 0;
- tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
- tu_cs_emit(cs, CP_WAIT_REG_MEM_0_FUNCTION(WRITE_EQ) |
- CP_WAIT_REG_MEM_0_POLL_MEMORY);
- tu_cs_emit_qw(cs, event->bo.iova); /* POLL_ADDR_LO/HI */
- tu_cs_emit(cs, CP_WAIT_REG_MEM_3_REF(1));
- tu_cs_emit(cs, CP_WAIT_REG_MEM_4_MASK(~0u));
- tu_cs_emit(cs, CP_WAIT_REG_MEM_5_DELAY_LOOP_CYCLES(20));
- }
+ tu_barrier(cmd, memoryBarrierCount, pMemoryBarriers,
+ bufferMemoryBarrierCount, pBufferMemoryBarriers,
+ imageMemoryBarrierCount, pImageMemoryBarriers, &info);
}
void
projects / mesa.git / blobdiff