#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)
tu_bo_list_add_info(struct tu_bo_list *list,
const struct drm_msm_gem_submit_bo *bo_info)
{
+ assert(bo_info->handle != 0);
+
for (uint32_t i = 0; i < list->count; ++i) {
if (list->bo_infos[i].handle == bo_info->handle) {
assert(list->bo_infos[i].presumed == bo_info->presumed);
return VK_SUCCESS;
}
-static VkResult
-tu_tiling_config_update_gmem_layout(struct tu_tiling_config *tiling,
- const struct tu_device *dev)
-{
- const uint32_t gmem_size = dev->physical_device->gmem_size;
- uint32_t offset = 0;
-
- for (uint32_t i = 0; i < tiling->buffer_count; i++) {
- /* 16KB-aligned */
- offset = align(offset, 0x4000);
-
- tiling->gmem_offsets[i] = offset;
- offset += tiling->tile0.extent.width * tiling->tile0.extent.height *
- tiling->buffer_cpp[i];
- }
-
- return offset <= gmem_size ? VK_SUCCESS : VK_ERROR_OUT_OF_DEVICE_MEMORY;
-}
-
static void
tu_tiling_config_update_tile_layout(struct tu_tiling_config *tiling,
- const struct tu_device *dev)
+ 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;
}
/* do not exceed gmem size */
- while (tu_tiling_config_update_gmem_layout(tiling, dev) != VK_SUCCESS) {
- if (tiling->tile0.extent.width > tiling->tile0.extent.height) {
+ 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(ra_width / tiling->tile_count.width, tile_align_w);
+ 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(ra_height / tiling->tile_count.height, tile_align_h);
+ align(DIV_ROUND_UP(ra_height, tiling->tile_count.height), tile_align_h);
}
}
}
const uint32_t used_pipe_count =
tiling->pipe_count.width * tiling->pipe_count.height;
const VkExtent2D last_pipe = {
- .width = tiling->tile_count.width % tiling->pipe0.width,
- .height = tiling->tile_count.height % tiling->pipe0.height,
+ .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);
sizeof(uint32_t) * (max_pipe_count - used_pipe_count));
}
-static void
-tu_tiling_config_update(struct tu_tiling_config *tiling,
- const struct tu_device *dev,
- const uint32_t *buffer_cpp,
- uint32_t buffer_count,
- const VkRect2D *render_area)
-{
- /* see if there is any real change */
- const bool ra_changed =
- render_area &&
- memcmp(&tiling->render_area, render_area, sizeof(*render_area));
- const bool buf_changed = tiling->buffer_count != buffer_count ||
- memcmp(tiling->buffer_cpp, buffer_cpp,
- sizeof(*buffer_cpp) * buffer_count);
- if (!ra_changed && !buf_changed)
- return;
-
- if (ra_changed)
- tiling->render_area = *render_area;
-
- if (buf_changed) {
- memcpy(tiling->buffer_cpp, buffer_cpp,
- sizeof(*buffer_cpp) * buffer_count);
- tiling->buffer_count = buffer_count;
- }
-
- tu_tiling_config_update_tile_layout(tiling, dev);
- tu_tiling_config_update_pipe_layout(tiling, dev);
- tu_tiling_config_update_pipes(tiling, dev);
-}
-
static void
tu_tiling_config_get_tile(const struct tu_tiling_config *tiling,
const struct tu_device *dev,
: tile->begin.y + tiling->tile0.extent.height;
}
-static enum a3xx_msaa_samples
-tu6_msaa_samples(uint32_t samples)
+enum a3xx_msaa_samples
+tu_msaa_samples(uint32_t samples)
{
switch (samples) {
case 1:
tu_cs_emit_write_reg(cs, cmd->marker_reg, ++cmd->marker_seqno);
}
-void
+unsigned
tu6_emit_event_write(struct tu_cmd_buffer *cmd,
struct tu_cs *cs,
enum vgt_event_type event,
bool need_seqno)
{
+ unsigned seqno = 0;
+
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);
- tu_cs_emit(cs, ++cmd->scratch_seqno);
+ seqno = ++cmd->scratch_seqno;
+ tu_cs_emit(cs, seqno);
}
+
+ return seqno;
}
static void
}
static void
-tu6_emit_zs(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+tu6_emit_flag_buffer(struct tu_cs *cs, const struct tu_image_view *iview)
+{
+ uint64_t va = tu_image_ubwc_base(iview->image, iview->base_mip, iview->base_layer);
+ uint32_t pitch = tu_image_ubwc_pitch(iview->image, iview->base_mip);
+ uint32_t size = tu_image_ubwc_size(iview->image, iview->base_mip);
+ if (iview->image->layout.ubwc_size) {
+ tu_cs_emit_qw(cs, va);
+ tu_cs_emit(cs, A6XX_RB_DEPTH_FLAG_BUFFER_PITCH_PITCH(pitch) |
+ A6XX_RB_DEPTH_FLAG_BUFFER_PITCH_ARRAY_PITCH(size >> 2));
+ } else {
+ tu_cs_emit_qw(cs, 0);
+ tu_cs_emit(cs, 0);
+ }
+}
+
+static void
+tu6_emit_zs(struct tu_cmd_buffer *cmd,
+ const struct tu_subpass *subpass,
+ struct tu_cs *cs)
{
- const struct tu_subpass *subpass = cmd->state.subpass;
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
const uint32_t a = subpass->depth_stencil_attachment.attachment;
if (a == VK_ATTACHMENT_UNUSED) {
return;
}
+ const struct tu_image_view *iview = fb->attachments[a].attachment;
+ enum a6xx_depth_format fmt = tu6_pipe2depth(iview->vk_format);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_BUFFER_INFO, 6);
+ tu_cs_emit(cs, A6XX_RB_DEPTH_BUFFER_INFO_DEPTH_FORMAT(fmt));
+ tu_cs_emit(cs, A6XX_RB_DEPTH_BUFFER_PITCH(tu_image_stride(iview->image, iview->base_mip)));
+ tu_cs_emit(cs, A6XX_RB_DEPTH_BUFFER_ARRAY_PITCH(iview->image->layout.layer_size));
+ tu_cs_emit_qw(cs, tu_image_base(iview->image, iview->base_mip, iview->base_layer));
+ tu_cs_emit(cs, cmd->state.pass->attachments[a].gmem_offset);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_SU_DEPTH_BUFFER_INFO, 1);
+ tu_cs_emit(cs, A6XX_GRAS_SU_DEPTH_BUFFER_INFO_DEPTH_FORMAT(fmt));
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_DEPTH_FLAG_BUFFER_BASE_LO, 3);
+ tu6_emit_flag_buffer(cs, iview);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_LRZ_BUFFER_BASE_LO, 5);
+ tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_FLAG_BUFFER_BASE_LO */
+ tu_cs_emit(cs, 0x00000000); /* RB_DEPTH_FLAG_BUFFER_BASE_HI */
+ tu_cs_emit(cs, 0x00000000); /* GRAS_LRZ_BUFFER_PITCH */
+ tu_cs_emit(cs, 0x00000000); /* GRAS_LRZ_FAST_CLEAR_BUFFER_BASE_LO */
+ tu_cs_emit(cs, 0x00000000); /* GRAS_LRZ_FAST_CLEAR_BUFFER_BASE_HI */
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_STENCIL_INFO, 1);
+ tu_cs_emit(cs, 0x00000000); /* RB_STENCIL_INFO */
+
/* enable zs? */
}
static void
-tu6_emit_mrt(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+tu6_emit_mrt(struct tu_cmd_buffer *cmd,
+ const struct tu_subpass *subpass,
+ struct tu_cs *cs)
{
const struct tu_framebuffer *fb = cmd->state.framebuffer;
- const struct tu_subpass *subpass = cmd->state.subpass;
- const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
unsigned char mrt_comp[MAX_RTS] = { 0 };
unsigned srgb_cntl = 0;
- uint32_t gmem_index = 0;
for (uint32_t i = 0; i < subpass->color_count; ++i) {
uint32_t a = subpass->color_attachments[i].attachment;
if (a == VK_ATTACHMENT_UNUSED)
continue;
const struct tu_image_view *iview = fb->attachments[a].attachment;
- const struct tu_image_level *slice =
- &iview->image->levels[iview->base_mip];
- const enum a6xx_tile_mode tile_mode = TILE6_LINEAR;
- uint32_t stride = 0;
- uint32_t offset = 0;
+ const enum a6xx_tile_mode tile_mode =
+ tu6_get_image_tile_mode(iview->image, iview->base_mip);
mrt_comp[i] = 0xf;
tu6_get_native_format(iview->vk_format);
assert(format && format->rb >= 0);
- offset = slice->offset + slice->size * iview->base_layer;
- stride = slice->pitch * vk_format_get_blocksize(iview->vk_format);
-
tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_BUF_INFO(i), 6);
tu_cs_emit(cs, A6XX_RB_MRT_BUF_INFO_COLOR_FORMAT(format->rb) |
A6XX_RB_MRT_BUF_INFO_COLOR_TILE_MODE(tile_mode) |
A6XX_RB_MRT_BUF_INFO_COLOR_SWAP(format->swap));
- tu_cs_emit(cs, A6XX_RB_MRT_PITCH(stride));
- tu_cs_emit(cs, A6XX_RB_MRT_ARRAY_PITCH(slice->size));
- tu_cs_emit_qw(cs, iview->image->bo->iova + iview->image->bo_offset +
- offset); /* BASE_LO/HI */
- tu_cs_emit(
- cs, tiling->gmem_offsets[gmem_index++]); /* RB_MRT[i].BASE_GMEM */
+ tu_cs_emit(cs, A6XX_RB_MRT_PITCH(tu_image_stride(iview->image, iview->base_mip)));
+ tu_cs_emit(cs, A6XX_RB_MRT_ARRAY_PITCH(iview->image->layout.layer_size));
+ tu_cs_emit_qw(cs, tu_image_base(iview->image, iview->base_mip, iview->base_layer));
+ tu_cs_emit(cs, cmd->state.pass->attachments[a].gmem_offset);
tu_cs_emit_pkt4(cs, REG_A6XX_SP_FS_MRT_REG(i), 1);
- tu_cs_emit(cs, A6XX_SP_FS_MRT_REG_COLOR_FORMAT(format->rb));
+ tu_cs_emit(cs, A6XX_SP_FS_MRT_REG_COLOR_FORMAT(format->rb) |
+ COND(vk_format_is_sint(iview->vk_format), A6XX_SP_FS_MRT_REG_COLOR_SINT) |
+ COND(vk_format_is_uint(iview->vk_format), A6XX_SP_FS_MRT_REG_COLOR_UINT));
-#if 0
- /* when we support UBWC, these would be the system memory
- * addr/pitch/etc:
- */
- tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_FLAG_BUFFER(i), 4);
- tu_cs_emit(cs, 0x00000000); /* RB_MRT_FLAG_BUFFER[i].ADDR_LO */
- tu_cs_emit(cs, 0x00000000); /* RB_MRT_FLAG_BUFFER[i].ADDR_HI */
- tu_cs_emit(cs, A6XX_RB_MRT_FLAG_BUFFER_PITCH(0));
- tu_cs_emit(cs, A6XX_RB_MRT_FLAG_BUFFER_ARRAY_PITCH(0));
-#endif
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_MRT_FLAG_BUFFER(i), 3);
+ tu6_emit_flag_buffer(cs, iview);
}
tu_cs_emit_pkt4(cs, REG_A6XX_RB_SRGB_CNTL, 1);
}
static void
-tu6_emit_msaa(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+tu6_emit_msaa(struct tu_cmd_buffer *cmd,
+ const struct tu_subpass *subpass,
+ struct tu_cs *cs)
{
- const struct tu_subpass *subpass = cmd->state.subpass;
- const enum a3xx_msaa_samples samples =
- tu6_msaa_samples(subpass->max_sample_count);
+ const enum a3xx_msaa_samples samples = tu_msaa_samples(subpass->samples);
tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_RAS_MSAA_CNTL, 2);
tu_cs_emit(cs, A6XX_SP_TP_RAS_MSAA_CNTL_SAMPLES(samples));
- tu_cs_emit(
- cs, A6XX_SP_TP_DEST_MSAA_CNTL_SAMPLES(samples) |
- ((samples == MSAA_ONE) ? A6XX_SP_TP_DEST_MSAA_CNTL_MSAA_DISABLE
- : 0));
+ tu_cs_emit(cs, A6XX_SP_TP_DEST_MSAA_CNTL_SAMPLES(samples) |
+ COND(samples == MSAA_ONE, A6XX_SP_TP_DEST_MSAA_CNTL_MSAA_DISABLE));
tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_RAS_MSAA_CNTL, 2);
tu_cs_emit(cs, A6XX_GRAS_RAS_MSAA_CNTL_SAMPLES(samples));
- tu_cs_emit(
- cs,
- A6XX_GRAS_DEST_MSAA_CNTL_SAMPLES(samples) |
- ((samples == MSAA_ONE) ? A6XX_GRAS_DEST_MSAA_CNTL_MSAA_DISABLE : 0));
+ tu_cs_emit(cs, A6XX_GRAS_DEST_MSAA_CNTL_SAMPLES(samples) |
+ COND(samples == MSAA_ONE, A6XX_GRAS_DEST_MSAA_CNTL_MSAA_DISABLE));
tu_cs_emit_pkt4(cs, REG_A6XX_RB_RAS_MSAA_CNTL, 2);
tu_cs_emit(cs, A6XX_RB_RAS_MSAA_CNTL_SAMPLES(samples));
- tu_cs_emit(
- cs,
- A6XX_RB_DEST_MSAA_CNTL_SAMPLES(samples) |
- ((samples == MSAA_ONE) ? A6XX_RB_DEST_MSAA_CNTL_MSAA_DISABLE : 0));
+ tu_cs_emit(cs, A6XX_RB_DEST_MSAA_CNTL_SAMPLES(samples) |
+ COND(samples == MSAA_ONE, A6XX_RB_DEST_MSAA_CNTL_MSAA_DISABLE));
tu_cs_emit_pkt4(cs, REG_A6XX_RB_MSAA_CNTL, 1);
tu_cs_emit(cs, A6XX_RB_MSAA_CNTL_SAMPLES(samples));
}
static void
-tu6_emit_blit_scissor(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+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 uint32_t x1 = render_area->offset.x;
- const uint32_t y1 = render_area->offset.y;
- const uint32_t x2 = x1 + render_area->extent.width - 1;
- const uint32_t y2 = y1 + render_area->extent.height - 1;
+ 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;
+ }
tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_SCISSOR_TL, 2);
tu_cs_emit(cs,
struct tu_cs *cs,
const struct tu_image_view *iview,
uint32_t gmem_offset,
- uint32_t blit_info)
+ bool resolve)
{
- const struct tu_image_level *slice =
- &iview->image->levels[iview->base_mip];
- const uint32_t offset = slice->offset + slice->size * iview->base_layer;
- const uint32_t stride =
- slice->pitch * vk_format_get_blocksize(iview->vk_format);
- const enum a6xx_tile_mode tile_mode = TILE6_LINEAR;
- const enum a3xx_msaa_samples samples = tu6_msaa_samples(1);
-
tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_INFO, 1);
- tu_cs_emit(cs, blit_info);
+ tu_cs_emit(cs, resolve ? 0 : (A6XX_RB_BLIT_INFO_UNK0 | A6XX_RB_BLIT_INFO_GMEM));
- /* tile mode? */
const struct tu_native_format *format =
tu6_get_native_format(iview->vk_format);
assert(format && format->rb >= 0);
+ enum a6xx_tile_mode tile_mode =
+ tu6_get_image_tile_mode(iview->image, iview->base_mip);
tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_DST_INFO, 5);
tu_cs_emit(cs, A6XX_RB_BLIT_DST_INFO_TILE_MODE(tile_mode) |
- A6XX_RB_BLIT_DST_INFO_SAMPLES(samples) |
- A6XX_RB_BLIT_DST_INFO_COLOR_FORMAT(format->rb) |
- A6XX_RB_BLIT_DST_INFO_COLOR_SWAP(format->swap));
- tu_cs_emit_qw(cs,
- iview->image->bo->iova + iview->image->bo_offset + offset);
- tu_cs_emit(cs, A6XX_RB_BLIT_DST_PITCH(stride));
- tu_cs_emit(cs, A6XX_RB_BLIT_DST_ARRAY_PITCH(slice->size));
-
- tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_BASE_GMEM, 1);
- tu_cs_emit(cs, gmem_offset);
-}
-
-static void
-tu6_emit_blit_clear(struct tu_cmd_buffer *cmd,
- struct tu_cs *cs,
- const struct tu_image_view *iview,
- uint32_t gmem_offset,
- const VkClearValue *clear_value)
-{
- const enum a6xx_tile_mode tile_mode = TILE6_LINEAR;
- const enum a3xx_msaa_samples samples = tu6_msaa_samples(1);
-
- const struct tu_native_format *format =
- tu6_get_native_format(iview->vk_format);
- assert(format && format->rb >= 0);
- /* must be WZYX; other values are ignored */
- const enum a3xx_color_swap swap = WZYX;
-
- tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_DST_INFO, 1);
- tu_cs_emit(cs, A6XX_RB_BLIT_DST_INFO_TILE_MODE(tile_mode) |
- A6XX_RB_BLIT_DST_INFO_SAMPLES(samples) |
+ A6XX_RB_BLIT_DST_INFO_SAMPLES(tu_msaa_samples(iview->image->samples)) |
A6XX_RB_BLIT_DST_INFO_COLOR_FORMAT(format->rb) |
- A6XX_RB_BLIT_DST_INFO_COLOR_SWAP(swap));
+ A6XX_RB_BLIT_DST_INFO_COLOR_SWAP(format->swap) |
+ COND(iview->image->layout.ubwc_size,
+ A6XX_RB_BLIT_DST_INFO_FLAGS));
+ tu_cs_emit_qw(cs, tu_image_base(iview->image, iview->base_mip, iview->base_layer));
+ tu_cs_emit(cs, A6XX_RB_BLIT_DST_PITCH(tu_image_stride(iview->image, iview->base_mip)));
+ tu_cs_emit(cs, A6XX_RB_BLIT_DST_ARRAY_PITCH(iview->image->layout.layer_size));
- tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_INFO, 1);
- tu_cs_emit(cs, A6XX_RB_BLIT_INFO_GMEM | A6XX_RB_BLIT_INFO_CLEAR_MASK(0xf));
+ if (iview->image->layout.ubwc_size) {
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_FLAG_DST_LO, 3);
+ tu6_emit_flag_buffer(cs, iview);
+ }
tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_BASE_GMEM, 1);
tu_cs_emit(cs, gmem_offset);
-
- tu_cs_emit_pkt4(cs, REG_A6XX_RB_UNKNOWN_88D0, 1);
- tu_cs_emit(cs, 0);
-
- /* pack clear_value into WZYX order */
- uint32_t clear_vals[4] = { 0 };
- tu_pack_clear_value(clear_value, iview->vk_format, clear_vals);
-
- tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_CLEAR_COLOR_DW0, 4);
- tu_cs_emit(cs, clear_vals[0]);
- tu_cs_emit(cs, clear_vals[1]);
- tu_cs_emit(cs, clear_vals[2]);
- tu_cs_emit(cs, clear_vals[3]);
}
static void
cs, A6XX_SP_TP_WINDOW_OFFSET_X(x1) | A6XX_SP_TP_WINDOW_OFFSET_Y(y1));
}
+static bool
+use_hw_binning(struct tu_cmd_buffer *cmd)
+{
+ const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+
+ if (unlikely(cmd->device->physical_device->instance->debug_flags & TU_DEBUG_NOBIN))
+ return false;
+
+ return (tiling->tile_count.width * tiling->tile_count.height) > 2;
+}
+
static void
tu6_emit_tile_select(struct tu_cmd_buffer *cmd,
struct tu_cs *cs,
const struct tu_tile *tile)
{
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
- tu_cs_emit(cs, A2XX_CP_SET_MARKER_0_MODE(0x7));
+ tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(0x7));
tu6_emit_marker(cmd, cs);
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
- tu_cs_emit(cs, A2XX_CP_SET_MARKER_0_MODE(RM6_GMEM) | 0x10);
+ tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_GMEM) | 0x10);
tu6_emit_marker(cmd, cs);
const uint32_t x1 = tile->begin.x;
tu_cs_emit_pkt4(cs, REG_A6XX_VPC_SO_OVERRIDE, 1);
tu_cs_emit(cs, A6XX_VPC_SO_OVERRIDE_SO_DISABLE);
- if (false) {
- /* hw binning? */
+ 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_COND_REG_EXEC, 2);
+ tu_cs_emit(cs, 0x10000000);
+ tu_cs_emit(cs, 11); /* conditionally execute next 11 dwords */
+
+ /* 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_MODE, 1);
+ tu_cs_emit(cs, 0x0);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_UNKNOWN_8804, 1);
+ tu_cs_emit(cs, 0x0);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_UNKNOWN_B304, 1);
+ tu_cs_emit(cs, 0x0);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_UNKNOWN_80A4, 1);
+ tu_cs_emit(cs, 0x0);
} else {
tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
tu_cs_emit(cs, 0x1);
}
static void
-tu6_emit_tile_load(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+tu6_emit_load_attachment(struct tu_cmd_buffer *cmd, struct tu_cs *cs, uint32_t a)
{
- const struct tu_framebuffer *fb = cmd->state.framebuffer;
- const struct tu_subpass *subpass = cmd->state.subpass;
const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
- const struct tu_attachment_state *attachments = cmd->state.attachments;
-
- tu6_emit_blit_scissor(cmd, cs);
-
- uint32_t gmem_index = 0;
- for (uint32_t i = 0; i < subpass->color_count; ++i) {
- const uint32_t a = subpass->color_attachments[i].attachment;
- if (a == VK_ATTACHMENT_UNUSED)
- continue;
+ 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];
- const struct tu_image_view *iview = fb->attachments[a].attachment;
- const struct tu_attachment_state *att = attachments + a;
- if (att->pending_clear_aspects) {
- assert(att->pending_clear_aspects == VK_IMAGE_ASPECT_COLOR_BIT);
- tu6_emit_blit_clear(cmd, cs, iview,
- tiling->gmem_offsets[gmem_index++],
- &att->clear_value);
- } else {
- tu6_emit_blit_info(cmd, cs, iview,
- tiling->gmem_offsets[gmem_index++],
- A6XX_RB_BLIT_INFO_UNK0 | A6XX_RB_BLIT_INFO_GMEM);
- }
+ 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);
}
-
- /* load/clear zs? */
}
static void
-tu6_emit_tile_store(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+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_tiling_config *tiling = &cmd->state.tiling_config;
+ 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;
- if (false) {
- /* hw binning? */
+ /* 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;
+
+ const struct tu_native_format *format =
+ tu6_get_native_format(iview->vk_format);
+ assert(format && format->rb >= 0);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_DST_INFO, 1);
+ tu_cs_emit(cs, A6XX_RB_BLIT_DST_INFO_COLOR_FORMAT(format->rb));
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_INFO, 1);
+ tu_cs_emit(cs, A6XX_RB_BLIT_INFO_GMEM | A6XX_RB_BLIT_INFO_CLEAR_MASK(clear_mask));
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_BASE_GMEM, 1);
+ tu_cs_emit(cs, attachment->gmem_offset);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_UNKNOWN_88D0, 1);
+ tu_cs_emit(cs, 0);
+
+ uint32_t clear_vals[4] = { 0 };
+ tu_pack_clear_value(&info->pClearValues[a], iview->vk_format, clear_vals);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_BLIT_CLEAR_COLOR_DW0, 4);
+ tu_cs_emit(cs, clear_vals[0]);
+ tu_cs_emit(cs, clear_vals[1]);
+ tu_cs_emit(cs, clear_vals[2]);
+ tu_cs_emit(cs, clear_vals[3]);
+
+ tu6_emit_blit(cmd, cs);
+}
+
+static void
+tu6_emit_store_attachment(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_blit_info(cmd, cs,
+ cmd->state.framebuffer->attachments[a].attachment,
+ cmd->state.pass->attachments[gmem_a].gmem_offset, true);
+ tu6_emit_blit(cmd, cs);
+}
+
+static void
+tu6_emit_tile_store(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+ const struct tu_render_pass *pass = cmd->state.pass;
+ const struct tu_subpass *subpass = &pass->subpasses[pass->subpass_count-1];
tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(0) |
tu6_emit_marker(cmd, cs);
tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
- tu_cs_emit(cs, A2XX_CP_SET_MARKER_0_MODE(RM6_RESOLVE) | 0x10);
+ tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_RESOLVE) | 0x10);
tu6_emit_marker(cmd, cs);
- tu6_emit_blit_scissor(cmd, cs);
+ tu6_emit_blit_scissor(cmd, cs, true);
- uint32_t gmem_index = 0;
- for (uint32_t i = 0; i < cmd->state.subpass->color_count; ++i) {
- uint32_t a = cmd->state.subpass->color_attachments[i].attachment;
- if (a == VK_ATTACHMENT_UNUSED)
- continue;
+ 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);
+ }
- const struct tu_image_view *iview = fb->attachments[a].attachment;
- tu6_emit_blit_info(cmd, cs, iview, tiling->gmem_offsets[gmem_index++],
- 0);
- tu6_emit_blit(cmd, cs);
+ 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_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_AB00, 0x5);
- tu_cs_emit_write_reg(cs, REG_A6XX_VFD_UNKNOWN_A009, 0x00000001);
+ 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_SRGB_CNTL, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8101, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8109, 0);
+ tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_SAMPLE_CNTL, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_GRAS_UNKNOWN_8110, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_RENDER_CONTROL0, 0x401);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_RENDER_CONTROL1, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_FS_OUTPUT_CNTL0, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8810, 0);
+ tu_cs_emit_write_reg(cs, REG_A6XX_RB_SAMPLE_CNTL, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8818, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_8819, 0);
tu_cs_emit_write_reg(cs, REG_A6XX_RB_UNKNOWN_881A, 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_9B06, 0);
- tu_cs_emit_write_reg(cs, REG_A6XX_PC_UNKNOWN_9B06, 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_A81B, 0);
tu_cs_sanity_check(cs);
}
+static void
+tu6_cache_flush(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+ unsigned seqno;
+
+ seqno = tu6_emit_event_write(cmd, cs, CACHE_FLUSH_AND_INV_EVENT, true);
+
+ 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));
+
+ seqno = tu6_emit_event_write(cmd, cs, CACHE_FLUSH_TS, true);
+
+ 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));
+}
+
+static void
+update_vsc_pipe(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
+{
+ const struct tu_tiling_config *tiling = &cmd->state.tiling_config;
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_VSC_BIN_SIZE, 3);
+ tu_cs_emit(cs, A6XX_VSC_BIN_SIZE_WIDTH(tiling->tile0.extent.width) |
+ A6XX_VSC_BIN_SIZE_HEIGHT(tiling->tile0.extent.height));
+ tu_cs_emit_qw(cs, cmd->vsc_data.iova + 32 * cmd->vsc_data_pitch); /* VSC_SIZE_ADDRESS_LO/HI */
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_VSC_BIN_COUNT, 1);
+ tu_cs_emit(cs, A6XX_VSC_BIN_COUNT_NX(tiling->tile_count.width) |
+ A6XX_VSC_BIN_COUNT_NY(tiling->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_pkt4(cs, REG_A6XX_VSC_PIPE_DATA2_ADDRESS_LO, 4);
+ tu_cs_emit_qw(cs, cmd->vsc_data2.iova);
+ tu_cs_emit(cs, cmd->vsc_data2_pitch);
+ tu_cs_emit(cs, cmd->vsc_data2.size);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_VSC_PIPE_DATA_ADDRESS_LO, 4);
+ tu_cs_emit_qw(cs, cmd->vsc_data.iova);
+ tu_cs_emit(cs, cmd->vsc_data_pitch);
+ tu_cs_emit(cs, cmd->vsc_data.size);
+}
+
+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 uint32_t used_pipe_count =
+ tiling->pipe_count.width * tiling->pipe_count.height;
+
+ /* Clear vsc_scratch: */
+ tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 3);
+ tu_cs_emit_qw(cs, cmd->scratch_bo.iova + 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_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_4_MASK(~0));
+ tu_cs_emit_qw(cs, cmd->scratch_bo.iova + VSC_SCRATCH);
+ tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(1 + cmd->vsc_data_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_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_4_MASK(~0));
+ tu_cs_emit_qw(cs, cmd->scratch_bo.iova + VSC_SCRATCH);
+ tu_cs_emit(cs, CP_COND_WRITE5_7_WRITE_DATA(3 + cmd->vsc_data2_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 + 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_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
+ tu_cs_emit(cs, 0x10000000);
+ tu_cs_emit(cs, 7); /* conditionally execute next 7 dwords */
+
+ /* 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 + 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)
+{
+ 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;
+
+ tu6_emit_window_scissor(cmd, cs, x1, y1, x2, y2);
+
+ tu6_emit_marker(cmd, cs);
+ tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
+ tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(RM6_BINNING));
+ tu6_emit_marker(cmd, cs);
+
+ tu_cs_emit_pkt7(cs, CP_SET_VISIBILITY_OVERRIDE, 1);
+ tu_cs_emit(cs, 0x1);
+
+ tu_cs_emit_pkt7(cs, CP_SET_MODE, 1);
+ tu_cs_emit(cs, 0x1);
+
+ tu_cs_emit_wfi(cs);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_VFD_MODE_CNTL, 1);
+ tu_cs_emit(cs, A6XX_VFD_MODE_CNTL_BINNING_PASS);
+
+ update_vsc_pipe(cmd, cs);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_PC_UNKNOWN_9805, 1);
+ tu_cs_emit(cs, 0x1);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_SP_UNKNOWN_A0F8, 1);
+ tu_cs_emit(cs, 0x1);
+
+ tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
+ tu_cs_emit(cs, UNK_2C);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_WINDOW_OFFSET, 1);
+ tu_cs_emit(cs, A6XX_RB_WINDOW_OFFSET_X(0) |
+ A6XX_RB_WINDOW_OFFSET_Y(0));
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_WINDOW_OFFSET, 1);
+ tu_cs_emit(cs, A6XX_SP_TP_WINDOW_OFFSET_X(0) |
+ A6XX_SP_TP_WINDOW_OFFSET_Y(0));
+
+ /* emit IB to binning drawcmds: */
+ tu_cs_emit_call(cs, &cmd->draw_cs);
+
+ tu_cs_emit_pkt7(cs, CP_SET_DRAW_STATE, 3);
+ tu_cs_emit(cs, CP_SET_DRAW_STATE__0_COUNT(0) |
+ CP_SET_DRAW_STATE__0_DISABLE_ALL_GROUPS |
+ CP_SET_DRAW_STATE__0_GROUP_ID(0));
+ tu_cs_emit(cs, CP_SET_DRAW_STATE__1_ADDR_LO(0));
+ tu_cs_emit(cs, CP_SET_DRAW_STATE__2_ADDR_HI(0));
+
+ 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);
+
+ tu_cs_emit_wfi(cs);
+
+ tu_cs_emit_pkt7(cs, CP_WAIT_FOR_ME, 0);
+
+ emit_vsc_overflow_test(cmd, cs);
+
+ 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_wfi(cs);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_RB_CCU_CNTL, 1);
+ tu_cs_emit(cs, 0x7c400004);
+
+ cmd->wait_for_idle = false;
+}
+
static void
tu6_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
- VkResult result = tu_cs_reserve_space(cmd->device, cs, 256);
+ VkResult result = tu_cs_reserve_space(cmd->device, cs, 1024);
if (result != VK_SUCCESS) {
cmd->record_result = result;
return;
tu_cs_emit_pkt4(cs, REG_A6XX_RB_CCU_CNTL, 1);
tu_cs_emit(cs, 0x7c400004); /* RB_CCU_CNTL */
- tu6_emit_zs(cmd, cs);
- tu6_emit_mrt(cmd, cs);
- tu6_emit_msaa(cmd, cs);
+ if (use_hw_binning(cmd)) {
+ tu6_emit_bin_size(cmd, cs, A6XX_RB_BIN_CONTROL_BINNING_PASS | 0x6000000);
+
+ tu6_emit_render_cntl(cmd, cs, true);
+
+ tu6_emit_binning_pass(cmd, cs);
+
+ tu6_emit_bin_size(cmd, cs, A6XX_RB_BIN_CONTROL_USE_VIZ | 0x6000000);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_VFD_MODE_CNTL, 1);
+ tu_cs_emit(cs, 0x0);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_PC_UNKNOWN_9805, 1);
+ tu_cs_emit(cs, 0x1);
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_SP_UNKNOWN_A0F8, 1);
+ tu_cs_emit(cs, 0x1);
- if (false) {
- /* hw binning? */
+ tu_cs_emit_pkt7(cs, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
+ tu_cs_emit(cs, 0x1);
} else {
tu6_emit_bin_size(cmd, cs, 0x6000000);
- /* no draws */
}
tu6_emit_render_cntl(cmd, cs, false);
struct tu_cs *cs,
const struct tu_tile *tile)
{
- const uint32_t render_tile_space = 64 + tu_cs_get_call_size(&cmd->draw_cs);
+ const uint32_t render_tile_space = 256 + tu_cs_get_call_size(&cmd->draw_cs);
VkResult result = tu_cs_reserve_space(cmd->device, cs, render_tile_space);
if (result != VK_SUCCESS) {
cmd->record_result = result;
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_emit_pkt7(cs, CP_COND_REG_EXEC, 2);
+ tu_cs_emit(cs, 0x10000000);
+ tu_cs_emit(cs, 2); /* conditionally execute next 2 dwords */
+
+ /* if (no overflow) */ {
+ tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
+ tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(0x5) | 0x10);
+ }
+ }
+
tu_cs_emit_ib(cs, &cmd->state.tile_store_ib);
tu_cs_sanity_check(cs);
}
tu_cs_emit_pkt4(cs, REG_A6XX_GRAS_LRZ_CNTL, 1);
- tu_cs_emit(cs, A6XX_GRAS_LRZ_CNTL_ENABLE | A6XX_GRAS_LRZ_CNTL_UNK3);
+ tu_cs_emit(cs, 0);
tu6_emit_lrz_flush(cmd, cs);
}
static void
-tu_cmd_prepare_tile_load_ib(struct tu_cmd_buffer *cmd)
+tu_cmd_prepare_tile_load_ib(struct tu_cmd_buffer *cmd,
+ const VkRenderPassBeginInfo *info)
{
- const uint32_t tile_load_space = 16 + 32 * MAX_RTS;
- const struct tu_subpass *subpass = cmd->state.subpass;
- struct tu_attachment_state *attachments = cmd->state.attachments;
+ const uint32_t tile_load_space =
+ 8 + (23+19) * cmd->state.pass->attachment_count +
+ 21 + (13 * cmd->state.subpass->color_count + 8) + 11;
+
struct tu_cs sub_cs;
- VkResult result = tu_cs_begin_sub_stream(cmd->device, &cmd->tile_cs,
+ VkResult result = tu_cs_begin_sub_stream(cmd->device, &cmd->sub_cs,
tile_load_space, &sub_cs);
if (result != VK_SUCCESS) {
cmd->record_result = result;
return;
}
- /* emit to tile-load sub_cs */
- tu6_emit_tile_load(cmd, &sub_cs);
+ tu6_emit_blit_scissor(cmd, &sub_cs, true);
- cmd->state.tile_load_ib = tu_cs_end_sub_stream(&cmd->tile_cs, &sub_cs);
+ for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
+ tu6_emit_load_attachment(cmd, &sub_cs, i);
- for (uint32_t i = 0; i < subpass->color_count; ++i) {
- const uint32_t a = subpass->color_attachments[i].attachment;
- if (a != VK_ATTACHMENT_UNUSED)
- attachments[a].pending_clear_aspects = 0;
- }
-}
+ tu6_emit_blit_scissor(cmd, &sub_cs, false);
-static void
-tu_cmd_prepare_tile_store_ib(struct tu_cmd_buffer *cmd)
-{
- const uint32_t tile_store_space = 32 + 32 * MAX_RTS;
+ for (uint32_t i = 0; i < cmd->state.pass->attachment_count; ++i)
+ tu6_emit_clear_attachment(cmd, &sub_cs, i, info);
+
+ /* 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, &sub_cs, CACHE_INVALIDATE, false);
+
+ tu6_emit_zs(cmd, cmd->state.subpass, &sub_cs);
+ tu6_emit_mrt(cmd, cmd->state.subpass, &sub_cs);
+ tu6_emit_msaa(cmd, cmd->state.subpass, &sub_cs);
+
+ cmd->state.tile_load_ib = tu_cs_end_sub_stream(&cmd->sub_cs, &sub_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;
struct tu_cs sub_cs;
- VkResult result = tu_cs_begin_sub_stream(cmd->device, &cmd->tile_cs,
+ VkResult result = tu_cs_begin_sub_stream(cmd->device, &cmd->sub_cs,
tile_store_space, &sub_cs);
if (result != VK_SUCCESS) {
cmd->record_result = result;
/* emit to tile-store sub_cs */
tu6_emit_tile_store(cmd, &sub_cs);
- cmd->state.tile_store_ib = tu_cs_end_sub_stream(&cmd->tile_cs, &sub_cs);
+ cmd->state.tile_store_ib = tu_cs_end_sub_stream(&cmd->sub_cs, &sub_cs);
}
static void
const VkRect2D *render_area)
{
const struct tu_device *dev = cmd->device;
- const struct tu_render_pass *pass = cmd->state.pass;
- const struct tu_subpass *subpass = cmd->state.subpass;
struct tu_tiling_config *tiling = &cmd->state.tiling_config;
- uint32_t buffer_cpp[MAX_RTS + 2];
- uint32_t buffer_count = 0;
-
- for (uint32_t i = 0; i < subpass->color_count; ++i) {
- const uint32_t a = subpass->color_attachments[i].attachment;
- if (a == VK_ATTACHMENT_UNUSED)
- continue;
-
- const struct tu_render_pass_attachment *att = &pass->attachments[a];
- buffer_cpp[buffer_count++] =
- vk_format_get_blocksize(att->format) * att->samples;
- }
-
- if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) {
- const uint32_t a = subpass->depth_stencil_attachment.attachment;
- const struct tu_render_pass_attachment *att = &pass->attachments[a];
-
- /* TODO */
- assert(att->format != VK_FORMAT_D32_SFLOAT_S8_UINT);
-
- buffer_cpp[buffer_count++] =
- vk_format_get_blocksize(att->format) * att->samples;
- }
+ tiling->render_area = *render_area;
- tu_tiling_config_update(tiling, dev, buffer_cpp, buffer_count,
- 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 = {
tu_bo_list_init(&cmd_buffer->bo_list);
tu_cs_init(&cmd_buffer->cs, TU_CS_MODE_GROW, 4096);
tu_cs_init(&cmd_buffer->draw_cs, TU_CS_MODE_GROW, 4096);
- tu_cs_init(&cmd_buffer->tile_cs, TU_CS_MODE_SUB_STREAM, 1024);
+ tu_cs_init(&cmd_buffer->sub_cs, TU_CS_MODE_SUB_STREAM, 2048);
*pCommandBuffer = tu_cmd_buffer_to_handle(cmd_buffer);
if (result != VK_SUCCESS)
return result;
+#define VSC_DATA_SIZE(pitch) ((pitch) * 32 + 0x100) /* extra size to store VSC_SIZE */
+#define VSC_DATA2_SIZE(pitch) ((pitch) * 32)
+
+ /* TODO: resize on overflow or compute a max size from # of vertices in renderpass?? */
+ cmd_buffer->vsc_data_pitch = 0x440 * 4;
+ cmd_buffer->vsc_data2_pitch = 0x1040 * 4;
+
+ result = tu_bo_init_new(device, &cmd_buffer->vsc_data, VSC_DATA_SIZE(cmd_buffer->vsc_data_pitch));
+ if (result != VK_SUCCESS)
+ goto fail_vsc_data;
+
+ result = tu_bo_init_new(device, &cmd_buffer->vsc_data2, VSC_DATA2_SIZE(cmd_buffer->vsc_data2_pitch));
+ if (result != VK_SUCCESS)
+ goto fail_vsc_data2;
+
return VK_SUCCESS;
+
+fail_vsc_data2:
+ tu_bo_finish(cmd_buffer->device, &cmd_buffer->vsc_data);
+fail_vsc_data:
+ tu_bo_finish(cmd_buffer->device, &cmd_buffer->scratch_bo);
+ return result;
}
static void
tu_cmd_buffer_destroy(struct tu_cmd_buffer *cmd_buffer)
{
tu_bo_finish(cmd_buffer->device, &cmd_buffer->scratch_bo);
+ tu_bo_finish(cmd_buffer->device, &cmd_buffer->vsc_data);
+ tu_bo_finish(cmd_buffer->device, &cmd_buffer->vsc_data2);
list_del(&cmd_buffer->pool_link);
tu_cs_finish(cmd_buffer->device, &cmd_buffer->cs);
tu_cs_finish(cmd_buffer->device, &cmd_buffer->draw_cs);
- tu_cs_finish(cmd_buffer->device, &cmd_buffer->tile_cs);
+ tu_cs_finish(cmd_buffer->device, &cmd_buffer->sub_cs);
tu_bo_list_destroy(&cmd_buffer->bo_list);
vk_free(&cmd_buffer->pool->alloc, cmd_buffer);
tu_bo_list_reset(&cmd_buffer->bo_list);
tu_cs_reset(cmd_buffer->device, &cmd_buffer->cs);
tu_cs_reset(cmd_buffer->device, &cmd_buffer->draw_cs);
- tu_cs_reset(cmd_buffer->device, &cmd_buffer->tile_cs);
+ tu_cs_reset(cmd_buffer->device, &cmd_buffer->sub_cs);
for (unsigned i = 0; i < VK_PIPELINE_BIND_POINT_RANGE_SIZE; i++) {
cmd_buffer->descriptors[i].dirty = 0;
return cmd_buffer->record_result;
}
-static VkResult
-tu_cmd_state_setup_attachments(struct tu_cmd_buffer *cmd_buffer,
- const VkRenderPassBeginInfo *info)
-{
- struct tu_cmd_state *state = &cmd_buffer->state;
- const struct tu_framebuffer *fb = state->framebuffer;
- const struct tu_render_pass *pass = state->pass;
-
- 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_buffer->bo_list, iview->image->bo,
- MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
- }
-
- if (pass->attachment_count == 0) {
- state->attachments = NULL;
- return VK_SUCCESS;
- }
-
- state->attachments =
- vk_alloc(&cmd_buffer->pool->alloc,
- pass->attachment_count * sizeof(state->attachments[0]), 8,
- VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
- if (state->attachments == NULL) {
- cmd_buffer->record_result = VK_ERROR_OUT_OF_HOST_MEMORY;
- return cmd_buffer->record_result;
- }
-
- for (uint32_t i = 0; i < pass->attachment_count; ++i) {
- const struct tu_render_pass_attachment *att = &pass->attachments[i];
- VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
- VkImageAspectFlags clear_aspects = 0;
-
- if (att_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
- /* color attachment */
- if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
- clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
- }
- } else {
- /* depthstencil attachment */
- if ((att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) &&
- att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
- clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
- if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
- att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_DONT_CARE)
- clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
- }
- if ((att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) &&
- att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
- clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
- }
- }
-
- state->attachments[i].pending_clear_aspects = clear_aspects;
- state->attachments[i].cleared_views = 0;
- if (clear_aspects && info) {
- assert(info->clearValueCount > i);
- state->attachments[i].clear_value = info->pClearValues[i];
- }
-
- state->attachments[i].current_layout = att->initial_layout;
- }
-
- return VK_SUCCESS;
-}
-
VkResult
tu_AllocateCommandBuffers(VkDevice _device,
const VkCommandBufferAllocateInfo *pAllocateInfo,
for (i = 0; i < pAllocateInfo->commandBufferCount; i++) {
- if (!list_empty(&pool->free_cmd_buffers)) {
+ if (!list_is_empty(&pool->free_cmd_buffers)) {
struct tu_cmd_buffer *cmd_buffer = list_first_entry(
&pool->free_cmd_buffers, struct tu_cmd_buffer, pool_link);
cmd_buffer->usage_flags = pBeginInfo->flags;
tu_cs_begin(&cmd_buffer->cs);
+ tu_cs_begin(&cmd_buffer->draw_cs);
cmd_buffer->marker_seqno = 0;
cmd_buffer->scratch_seqno = 0;
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];
}
cmd_buffer->status = TU_CMD_BUFFER_STATUS_RECORDING;
uint32_t dynamicOffsetCount,
const uint32_t *pDynamicOffsets)
{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, 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);
+
+ 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;
+ assert(dyn_idx < dynamicOffsetCount);
+
+ descriptors_state->dynamic_buffers[idx] =
+ set->dynamic_descriptors[j].va + pDynamicOffsets[dyn_idx];
+ }
+ }
+
+ cmd_buffer->state.dirty |= TU_CMD_DIRTY_DESCRIPTOR_SETS;
}
void
uint32_t size,
const void *pValues)
{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ memcpy((void*) cmd->push_constants + offset, pValues, size);
+ cmd->state.dirty |= TU_CMD_DIRTY_PUSH_CONSTANTS;
}
VkResult
MSM_SUBMIT_BO_WRITE);
}
+ 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);
+ }
+
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],
MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
}
- for (uint32_t i = 0; i < cmd_buffer->tile_cs.bo_count; i++) {
- tu_bo_list_add(&cmd_buffer->bo_list, cmd_buffer->tile_cs.bos[i],
+ for (uint32_t i = 0; i < cmd_buffer->sub_cs.bo_count; i++) {
+ tu_bo_list_add(&cmd_buffer->bo_list, cmd_buffer->sub_cs.bos[i],
MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_DUMP);
}
tu_cs_end(&cmd_buffer->cs);
-
- assert(!cmd_buffer->state.attachments);
+ tu_cs_end(&cmd_buffer->draw_cs);
cmd_buffer->status = TU_CMD_BUFFER_STATUS_EXECUTABLE;
cmd->state.dirty |= TU_CMD_DIRTY_PIPELINE;
break;
case VK_PIPELINE_BIND_POINT_COMPUTE:
- tu_finishme("binding compute pipeline");
+ 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);
+ }
}
void
uint32_t commandBufferCount,
const VkCommandBuffer *pCmdBuffers)
{
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ VkResult result;
+
+ assert(commandBufferCount > 0);
+
+ for (uint32_t i = 0; i < commandBufferCount; i++) {
+ TU_FROM_HANDLE(tu_cmd_buffer, secondary, pCmdBuffers[i]);
+
+ result = tu_bo_list_merge(&cmd->bo_list, &secondary->bo_list);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ break;
+ }
+
+ result = tu_cs_add_entries(&cmd->draw_cs, &secondary->draw_cs);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ break;
+ }
+ }
+ cmd->state.dirty = ~0u; /* TODO: set dirty only what needs to be */
}
VkResult
void
tu_TrimCommandPool(VkDevice device,
VkCommandPool commandPool,
- VkCommandPoolTrimFlagsKHR flags)
+ VkCommandPoolTrimFlags flags)
{
TU_FROM_HANDLE(tu_cmd_pool, pool, commandPool);
const VkRenderPassBeginInfo *pRenderPassBegin,
VkSubpassContents contents)
{
- TU_FROM_HANDLE(tu_cmd_buffer, cmd_buffer, commandBuffer);
+ TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
TU_FROM_HANDLE(tu_render_pass, pass, pRenderPassBegin->renderPass);
- TU_FROM_HANDLE(tu_framebuffer, framebuffer, pRenderPassBegin->framebuffer);
- VkResult result;
+ TU_FROM_HANDLE(tu_framebuffer, fb, pRenderPassBegin->framebuffer);
- cmd_buffer->state.pass = pass;
- cmd_buffer->state.subpass = pass->subpasses;
- cmd_buffer->state.framebuffer = framebuffer;
+ cmd->state.pass = pass;
+ cmd->state.subpass = pass->subpasses;
+ cmd->state.framebuffer = fb;
- result = tu_cmd_state_setup_attachments(cmd_buffer, pRenderPassBegin);
- if (result != VK_SUCCESS)
- return;
+ tu_cmd_update_tiling_config(cmd, &pRenderPassBegin->renderArea);
+ tu_cmd_prepare_tile_load_ib(cmd, pRenderPassBegin);
+ tu_cmd_prepare_tile_store_ib(cmd);
- tu_cmd_update_tiling_config(cmd_buffer, &pRenderPassBegin->renderArea);
- tu_cmd_prepare_tile_load_ib(cmd_buffer);
- tu_cmd_prepare_tile_store_ib(cmd_buffer);
+ /* note: use_hw_binning only checks tiling config */
+ if (use_hw_binning(cmd))
+ cmd->use_vsc_data = true;
- /* draw_cs should contain entries only for this render pass */
- assert(!cmd_buffer->draw_cs.entry_count);
- tu_cs_begin(&cmd_buffer->draw_cs);
+ 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);
+ }
}
void
-tu_CmdBeginRenderPass2KHR(VkCommandBuffer commandBuffer,
- const VkRenderPassBeginInfo *pRenderPassBeginInfo,
- const VkSubpassBeginInfoKHR *pSubpassBeginInfo)
+tu_CmdBeginRenderPass2(VkCommandBuffer commandBuffer,
+ const VkRenderPassBeginInfo *pRenderPassBeginInfo,
+ const VkSubpassBeginInfoKHR *pSubpassBeginInfo)
{
tu_CmdBeginRenderPass(commandBuffer, pRenderPassBeginInfo,
pSubpassBeginInfo->contents);
tu_CmdNextSubpass(VkCommandBuffer commandBuffer, VkSubpassContents contents)
{
TU_FROM_HANDLE(tu_cmd_buffer, cmd, commandBuffer);
+ const struct tu_render_pass *pass = cmd->state.pass;
+ struct tu_cs *cs = &cmd->draw_cs;
+
+ VkResult result = tu_cs_reserve_space(cmd->device, cs, 1024);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ return;
+ }
+
+ const struct tu_subpass *subpass = cmd->state.subpass++;
+ /* TODO:
+ * if msaa samples change between subpasses,
+ * attachment store is broken for some attachments
+ */
+ 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_store_attachment(cmd, cs, a,
+ subpass->color_attachments[i].attachment);
+ }
+ }
+ }
- tu_cmd_render_tiles(cmd);
+ /* 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);
- cmd->state.subpass++;
+ /* emit mrt/zs/msaa 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);
- tu_cmd_update_tiling_config(cmd, NULL);
- tu_cmd_prepare_tile_load_ib(cmd);
- tu_cmd_prepare_tile_store_ib(cmd);
+ /* 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;
+ const struct tu_image_view *iview =
+ cmd->state.framebuffer->attachments[a].attachment;
+ if (a != VK_ATTACHMENT_UNUSED && pass->attachments[a].gmem_offset >= 0) {
+ tu_finishme("missing GMEM->GMEM resolve, performance will suffer\n");
+ tu6_emit_blit_info(cmd, cs, iview, pass->attachments[a].gmem_offset, false);
+ tu6_emit_blit(cmd, cs);
+ }
+ }
+ }
}
void
-tu_CmdNextSubpass2KHR(VkCommandBuffer commandBuffer,
- const VkSubpassBeginInfoKHR *pSubpassBeginInfo,
- const VkSubpassEndInfoKHR *pSubpassEndInfo)
+tu_CmdNextSubpass2(VkCommandBuffer commandBuffer,
+ const VkSubpassBeginInfoKHR *pSubpassBeginInfo,
+ const VkSubpassEndInfoKHR *pSubpassEndInfo)
{
tu_CmdNextSubpass(commandBuffer, pSubpassBeginInfo->contents);
}
uint64_t count_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_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,
+ TU_DRAW_STATE_FS_IBO,
+ TU_DRAW_STATE_VS_PARAMS,
TU_DRAW_STATE_COUNT,
};
{
enum tu_draw_state_group_id id;
uint32_t enable_mask;
- const struct tu_cs_entry *ib;
+ struct tu_cs_entry ib;
};
+const static struct tu_sampler*
+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 struct tu_sampler *immutable_samplers =
+ tu_immutable_samplers(set->layout, layout);
+
+ return &immutable_samplers[array_index];
+ }
+
+ switch (layout->type) {
+ case VK_DESCRIPTOR_TYPE_SAMPLER:
+ return (struct tu_sampler*) &set->mapped_ptr[layout->offset / 4];
+ case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+ return (struct tu_sampler*) &set->mapped_ptr[layout->offset / 4 + A6XX_TEX_CONST_DWORDS +
+ array_index *
+ (A6XX_TEX_CONST_DWORDS +
+ sizeof(struct tu_sampler) / 4)];
+ 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)
+{
+ 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 +
+ sizeof(struct tu_sampler) / 4)],
+ A6XX_TEX_CONST_DWORDS * 4);
+ break;
+ default:
+ unreachable("unimplemented descriptor type");
+ break;
+ }
+
+ if (layout->type == VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT) {
+ 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] = 0x100000 + 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 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,
+ gl_shader_stage type,
+ uint32_t *push_constants)
+{
+ 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;
+ }
+
+ /* 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);
+ }
+ }
+}
+
+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);
+
+ if (!num)
+ return;
+
+ 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++;
+ }
+ }
+
+ for (; emitted < anum; emitted++) {
+ tu_cs_emit(cs, 0xffffffff);
+ tu_cs_emit(cs, 0xffffffff);
+ }
+}
+
+static struct tu_cs_entry
+tu6_emit_consts(struct tu_cmd_buffer *cmd,
+ const struct tu_pipeline *pipeline,
+ struct tu_descriptor_state *descriptors_state,
+ gl_shader_stage type)
+{
+ struct tu_cs cs;
+ tu_cs_begin_sub_stream(cmd->device, &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->device, &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 *needs_border)
+{
+ struct tu_device *device = cmd->device;
+ 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(device, 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);
+ }
+ }
+
+ /* allocate and fill sampler state */
+ struct ts_cs_memory tex_samp = { 0 };
+ if (link->sampler_map.num_desc) {
+ result = tu_cs_alloc(device, 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 struct tu_sampler *sampler = sampler_ptr(descriptors_state,
+ &link->sampler_map,
+ i, j);
+ memcpy(&tex_samp.map[A6XX_TEX_SAMP_DWORDS * sampler_index++],
+ sampler->state, sizeof(sampler->state));
+ *needs_border |= sampler->needs_border;
+ }
+ }
+ }
+
+ 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(device, 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;
+}
+
+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_device *device = cmd->device;
+ struct tu_cs *draw_state = &cmd->sub_cs;
+ const struct tu_program_descriptor_linkage *link =
+ &pipeline->program.link[type];
+ VkResult result;
+
+ if (link->image_mapping.num_ibo == 0) {
+ *entry = (struct tu_cs_entry) {};
+ return VK_SUCCESS;
+ }
+
+ struct ts_cs_memory ibo_const;
+ result = tu_cs_alloc(device, draw_state, link->image_mapping.num_ibo,
+ A6XX_TEX_CONST_DWORDS, &ibo_const);
+ if (result != VK_SUCCESS)
+ return result;
+
+ for (unsigned i = 0; i < link->image_mapping.num_ibo; i++) {
+ unsigned idx = link->image_mapping.ibo_to_image[i];
+ uint32_t *dst = &ibo_const.map[A6XX_TEX_CONST_DWORDS * i];
+
+ if (idx & IBO_SSBO) {
+ idx &= ~IBO_SSBO;
+
+ uint64_t va = buffer_ptr(descriptors_state, &link->ssbo_map, idx,
+ 0 /* XXX */);
+ /* We don't expose robustBufferAccess, so leave the size unlimited. */
+ uint32_t sz = MAX_STORAGE_BUFFER_RANGE / 4;
+
+ dst[0] = A6XX_IBO_0_FMT(TFMT6_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;
+ } else {
+ tu_finishme("Emit images");
+ }
+ }
+
+ struct tu_cs cs;
+ result = tu_cs_begin_sub_stream(device, draw_state, 7, &cs);
+ if (result != VK_SUCCESS)
+ return result;
+
+ uint32_t opcode, ibo_addr_reg;
+ enum a6xx_state_block sb;
+ enum a6xx_state_type st;
+
+ switch (type) {
+ case MESA_SHADER_FRAGMENT:
+ opcode = CP_LOAD_STATE6;
+ st = ST6_SHADER;
+ sb = SB6_IBO;
+ ibo_addr_reg = REG_A6XX_SP_IBO_LO;
+ 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;
+ break;
+ default:
+ unreachable("unsupported stage for ibos");
+ }
+
+ /* 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(link->image_mapping.num_ibo));
+ 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;
+}
+
+struct PACKED bcolor_entry {
+ uint32_t fp32[4];
+ uint16_t ui16[4];
+ int16_t si16[4];
+ uint16_t fp16[4];
+ uint16_t rgb565;
+ uint16_t rgb5a1;
+ uint16_t rgba4;
+ uint8_t __pad0[2];
+ uint8_t ui8[4];
+ int8_t si8[4];
+ uint32_t rgb10a2;
+ uint32_t z24; /* also s8? */
+ uint16_t srgb[4]; /* appears to duplicate fp16[], but clamped, used for srgb */
+ uint8_t __pad1[56];
+} border_color[] = {
+ [VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK] = {},
+ [VK_BORDER_COLOR_INT_TRANSPARENT_BLACK] = {},
+ [VK_BORDER_COLOR_FLOAT_OPAQUE_BLACK] = {
+ .fp32[3] = 0x3f800000,
+ .ui16[3] = 0xffff,
+ .si16[3] = 0x7fff,
+ .fp16[3] = 0x3c00,
+ .rgb5a1 = 0x8000,
+ .rgba4 = 0xf000,
+ .ui8[3] = 0xff,
+ .si8[3] = 0x7f,
+ .rgb10a2 = 0xc0000000,
+ .srgb[3] = 0x3c00,
+ },
+ [VK_BORDER_COLOR_INT_OPAQUE_BLACK] = {
+ .fp32[3] = 1,
+ .fp16[3] = 1,
+ },
+ [VK_BORDER_COLOR_FLOAT_OPAQUE_WHITE] = {
+ .fp32[0 ... 3] = 0x3f800000,
+ .ui16[0 ... 3] = 0xffff,
+ .si16[0 ... 3] = 0x7fff,
+ .fp16[0 ... 3] = 0x3c00,
+ .rgb565 = 0xffff,
+ .rgb5a1 = 0xffff,
+ .rgba4 = 0xffff,
+ .ui8[0 ... 3] = 0xff,
+ .si8[0 ... 3] = 0x7f,
+ .rgb10a2 = 0xffffffff,
+ .z24 = 0xffffff,
+ .srgb[0 ... 3] = 0x3c00,
+ },
+ [VK_BORDER_COLOR_INT_OPAQUE_WHITE] = {
+ .fp32[0 ... 3] = 1,
+ .fp16[0 ... 3] = 1,
+ },
+};
+
+static VkResult
+tu6_emit_border_color(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs)
+{
+ STATIC_ASSERT(sizeof(struct bcolor_entry) == 128);
+
+ const struct tu_pipeline *pipeline = cmd->state.pipeline;
+ struct tu_descriptor_state *descriptors_state =
+ &cmd->descriptors[VK_PIPELINE_BIND_POINT_GRAPHICS];
+ const struct tu_descriptor_map *vs_sampler =
+ &pipeline->program.link[MESA_SHADER_VERTEX].sampler_map;
+ const struct tu_descriptor_map *fs_sampler =
+ &pipeline->program.link[MESA_SHADER_FRAGMENT].sampler_map;
+ struct ts_cs_memory ptr;
+
+ VkResult result = tu_cs_alloc(cmd->device, &cmd->sub_cs,
+ vs_sampler->num_desc + fs_sampler->num_desc,
+ 128 / 4,
+ &ptr);
+ if (result != VK_SUCCESS)
+ return result;
+
+ for (unsigned i = 0; i < vs_sampler->num; i++) {
+ for (unsigned j = 0; j < vs_sampler->array_size[i]; j++) {
+ const struct tu_sampler *sampler = sampler_ptr(descriptors_state,
+ vs_sampler, i, j);
+ memcpy(ptr.map, &border_color[sampler->border], 128);
+ ptr.map += 128 / 4;
+ }
+ }
+
+ for (unsigned i = 0; i < fs_sampler->num; i++) {
+ for (unsigned j = 0; j < fs_sampler->array_size[i]; j++) {
+ const struct tu_sampler *sampler = sampler_ptr(descriptors_state,
+ fs_sampler, i, j);
+ memcpy(ptr.map, &border_color[sampler->border], 128);
+ ptr.map += 128 / 4;
+ }
+ }
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_SP_TP_BORDER_COLOR_BASE_ADDR_LO, 2);
+ tu_cs_emit_qw(cs, ptr.iova);
+ return VK_SUCCESS;
+}
+
+static VkResult
tu6_bind_draw_states(struct tu_cmd_buffer *cmd,
struct tu_cs *cs,
const struct tu_draw_info *draw)
struct tu_draw_state_group draw_state_groups[TU_DRAW_STATE_COUNT];
uint32_t draw_state_group_count = 0;
+ struct tu_descriptor_state *descriptors_state =
+ &cmd->descriptors[VK_PIPELINE_BIND_POINT_GRAPHICS];
+
VkResult result = tu_cs_reserve_space(cmd->device, cs, 256);
- if (result != VK_SUCCESS) {
- cmd->record_result = result;
- return;
- }
+ if (result != VK_SUCCESS)
+ return result;
/* TODO lrz */
}
}
- /* TODO shader consts */
-
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 = 0x6,
- .ib = &pipeline->program.state_ib,
+ .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 = 0x1,
- .ib = &pipeline->program.binning_state_ib,
+ .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 = 0x6,
- .ib = &pipeline->vi.state_ib,
+ .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 = 0x1,
- .ib = &pipeline->vi.binning_state_ib,
+ .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 = 0x7,
- .ib = &pipeline->vp.state_ib,
+ .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 = 0x7,
- .ib = &pipeline->rast.state_ib,
+ .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 = 0x7,
- .ib = &pipeline->ds.state_ib,
+ .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 = 0x7,
- .ib = &pipeline->blend.state_ib,
+ .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_PIPELINE | TU_CMD_DIRTY_DESCRIPTOR_SETS)) {
+ bool needs_border = false;
+ struct tu_cs_entry vs_tex, fs_tex, fs_ibo;
+
+ result = tu6_emit_textures(cmd, pipeline, descriptors_state,
+ MESA_SHADER_VERTEX, &vs_tex, &needs_border);
+ if (result != VK_SUCCESS)
+ return result;
+
+ result = tu6_emit_textures(cmd, pipeline, descriptors_state,
+ MESA_SHADER_FRAGMENT, &fs_tex, &needs_border);
+ if (result != VK_SUCCESS)
+ return result;
+
+ result = tu6_emit_ibo(cmd, pipeline, descriptors_state,
+ MESA_SHADER_FRAGMENT, &fs_ibo);
+ if (result != VK_SUCCESS)
+ return result;
+
+ 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,
+ .enable_mask = ENABLE_DRAW,
+ .ib = fs_tex,
+ };
+ 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,
+ };
+
+ if (needs_border) {
+ result = tu6_emit_border_color(cmd, cs);
+ if (result != VK_SUCCESS)
+ return result;
+ }
+ }
+
+ 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) |
- CP_SET_DRAW_STATE__0_ENABLE_MASK(group->enable_mask) |
+ 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;
+ 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;
tu_cs_sanity_check(cs);
/* track BOs */
- if (cmd->state.dirty & TU_CMD_DIRTY_PIPELINE) {
- 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 (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];
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);
+ }
+ }
+ }
- cmd->state.dirty = 0;
+ /* Fragment shader state overwrites compute shader state, so flag the
+ * compute pipeline for re-emit.
+ */
+ cmd->state.dirty = TU_CMD_DIRTY_COMPUTE_PIPELINE;
+ return VK_SUCCESS;
}
static void
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(IGNORE_VISIBILITY) | 0x2000;
+ CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) | 0x2000;
tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 7);
tu_cs_emit(cs, cp_draw_indx);
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(IGNORE_VISIBILITY) | 0x2000;
+ CP_DRAW_INDX_OFFSET_0_VIS_CULL(USE_VISIBILITY) | 0x2000;
tu_cs_emit_pkt7(cs, CP_DRAW_INDX_OFFSET, 3);
tu_cs_emit(cs, cp_draw_indx);
tu_draw(struct tu_cmd_buffer *cmd, const struct tu_draw_info *draw)
{
struct tu_cs *cs = &cmd->draw_cs;
+ VkResult result;
- tu6_bind_draw_states(cmd, cs, draw);
+ result = tu6_bind_draw_states(cmd, cs, draw);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ return;
+ }
- VkResult result = tu_cs_reserve_space(cmd->device, cs, 32);
+ result = tu_cs_reserve_space(cmd->device, cs, 32);
if (result != VK_SUCCESS) {
cmd->record_result = result;
return;
}
/* TODO tu6_emit_marker should pick different regs depending on cs */
+
tu6_emit_marker(cmd, cs);
tu6_emit_draw_direct(cmd, cs, draw);
tu6_emit_marker(cmd, cs);
};
static void
-tu_dispatch(struct tu_cmd_buffer *cmd_buffer,
+tu_emit_compute_driver_params(struct tu_cs *cs, struct tu_pipeline *pipeline,
+ const struct tu_dispatch_info *info)
+{
+ gl_shader_stage type = MESA_SHADER_COMPUTE;
+ const struct tu_program_descriptor_linkage *link =
+ &pipeline->program.link[type];
+ const struct ir3_const_state *const_state = &link->const_state;
+ uint32_t offset = const_state->offsets.driver_param;
+
+ if (link->constlen <= offset)
+ return;
+
+ if (!info->indirect) {
+ uint32_t driver_params[IR3_DP_CS_COUNT] = {
+ [IR3_DP_NUM_WORK_GROUPS_X] = info->blocks[0],
+ [IR3_DP_NUM_WORK_GROUPS_Y] = info->blocks[1],
+ [IR3_DP_NUM_WORK_GROUPS_Z] = info->blocks[2],
+ [IR3_DP_LOCAL_GROUP_SIZE_X] = pipeline->compute.local_size[0],
+ [IR3_DP_LOCAL_GROUP_SIZE_Y] = pipeline->compute.local_size[1],
+ [IR3_DP_LOCAL_GROUP_SIZE_Z] = pipeline->compute.local_size[2],
+ };
+
+ uint32_t num_consts = MIN2(const_state->num_driver_params,
+ (link->constlen - offset) * 4);
+ /* push constants */
+ tu_cs_emit_pkt7(cs, tu6_stage2opcode(type), 3 + num_consts);
+ 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_consts / 4));
+ tu_cs_emit(cs, 0);
+ tu_cs_emit(cs, 0);
+ uint32_t i;
+ for (i = 0; i < num_consts; i++)
+ tu_cs_emit(cs, driver_params[i]);
+ } else {
+ tu_finishme("Indirect driver params");
+ }
+}
+
+static void
+tu_dispatch(struct tu_cmd_buffer *cmd,
const struct tu_dispatch_info *info)
{
+ struct tu_cs *cs = &cmd->cs;
+ struct tu_pipeline *pipeline = cmd->state.compute_pipeline;
+ struct tu_descriptor_state *descriptors_state =
+ &cmd->descriptors[VK_PIPELINE_BIND_POINT_COMPUTE];
+
+ VkResult result = tu_cs_reserve_space(cmd->device, cs, 256);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ return;
+ }
+
+ if (cmd->state.dirty & TU_CMD_DIRTY_COMPUTE_PIPELINE)
+ tu_cs_emit_ib(cs, &pipeline->program.state_ib);
+
+ struct tu_cs_entry ib;
+
+ ib = tu6_emit_consts(cmd, pipeline, descriptors_state, MESA_SHADER_COMPUTE);
+ if (ib.size)
+ tu_cs_emit_ib(cs, &ib);
+
+ tu_emit_compute_driver_params(cs, pipeline, info);
+
+ bool needs_border;
+ result = tu6_emit_textures(cmd, pipeline, descriptors_state,
+ MESA_SHADER_COMPUTE, &ib, &needs_border);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ return;
+ }
+
+ if (ib.size)
+ tu_cs_emit_ib(cs, &ib);
+
+ if (needs_border)
+ tu_finishme("compute border color");
+
+ 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);
+ }
+ }
+ }
+
+ /* Compute shader state overwrites fragment shader state, so we flag the
+ * graphics pipeline for re-emit.
+ */
+ cmd->state.dirty = TU_CMD_DIRTY_PIPELINE;
+
+ tu_cs_emit_pkt7(cs, CP_SET_MARKER, 1);
+ tu_cs_emit(cs, A6XX_CP_SET_MARKER_0_MODE(0x8));
+
+ const uint32_t *local_size = pipeline->compute.local_size;
+ const uint32_t *num_groups = info->blocks;
+ tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CS_NDRANGE_0, 7);
+ tu_cs_emit(cs,
+ A6XX_HLSQ_CS_NDRANGE_0_KERNELDIM(3) |
+ A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEX(local_size[0] - 1) |
+ A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEY(local_size[1] - 1) |
+ A6XX_HLSQ_CS_NDRANGE_0_LOCALSIZEZ(local_size[2] - 1));
+ tu_cs_emit(cs, A6XX_HLSQ_CS_NDRANGE_1_GLOBALSIZE_X(local_size[0] * num_groups[0]));
+ tu_cs_emit(cs, 0); /* HLSQ_CS_NDRANGE_2_GLOBALOFF_X */
+ tu_cs_emit(cs, A6XX_HLSQ_CS_NDRANGE_3_GLOBALSIZE_Y(local_size[1] * num_groups[1]));
+ tu_cs_emit(cs, 0); /* HLSQ_CS_NDRANGE_4_GLOBALOFF_Y */
+ tu_cs_emit(cs, A6XX_HLSQ_CS_NDRANGE_5_GLOBALSIZE_Z(local_size[2] * num_groups[2]));
+ tu_cs_emit(cs, 0); /* HLSQ_CS_NDRANGE_6_GLOBALOFF_Z */
+
+ tu_cs_emit_pkt4(cs, REG_A6XX_HLSQ_CS_KERNEL_GROUP_X, 3);
+ tu_cs_emit(cs, 1); /* HLSQ_CS_KERNEL_GROUP_X */
+ tu_cs_emit(cs, 1); /* HLSQ_CS_KERNEL_GROUP_Y */
+ tu_cs_emit(cs, 1); /* HLSQ_CS_KERNEL_GROUP_Z */
+
+ if (info->indirect) {
+ uint64_t iova = tu_buffer_iova(info->indirect) + info->indirect_offset;
+
+ tu_bo_list_add(&cmd->bo_list, info->indirect->bo,
+ MSM_SUBMIT_BO_READ | MSM_SUBMIT_BO_WRITE);
+
+ tu_cs_emit_pkt7(cs, CP_EXEC_CS_INDIRECT, 4);
+ tu_cs_emit(cs, 0x00000000);
+ tu_cs_emit_qw(cs, iova);
+ tu_cs_emit(cs,
+ A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEX(local_size[0] - 1) |
+ A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEY(local_size[1] - 1) |
+ A5XX_CP_EXEC_CS_INDIRECT_3_LOCALSIZEZ(local_size[2] - 1));
+ } else {
+ tu_cs_emit_pkt7(cs, CP_EXEC_CS, 4);
+ tu_cs_emit(cs, 0x00000000);
+ tu_cs_emit(cs, CP_EXEC_CS_1_NGROUPS_X(info->blocks[0]));
+ tu_cs_emit(cs, CP_EXEC_CS_2_NGROUPS_Y(info->blocks[1]));
+ tu_cs_emit(cs, CP_EXEC_CS_3_NGROUPS_Z(info->blocks[2]));
+ }
+
+ tu_cs_emit_wfi(cs);
+
+ tu6_emit_cache_flush(cmd, cs);
}
void
/* discard draw_cs entries now that the tiles are rendered */
tu_cs_discard_entries(&cmd_buffer->draw_cs);
-
- vk_free(&cmd_buffer->pool->alloc, cmd_buffer->state.attachments);
- cmd_buffer->state.attachments = NULL;
+ tu_cs_begin(&cmd_buffer->draw_cs);
cmd_buffer->state.pass = NULL;
cmd_buffer->state.subpass = NULL;
}
void
-tu_CmdEndRenderPass2KHR(VkCommandBuffer commandBuffer,
- const VkSubpassEndInfoKHR *pSubpassEndInfo)
+tu_CmdEndRenderPass2(VkCommandBuffer commandBuffer,
+ const VkSubpassEndInfoKHR *pSubpassEndInfo)
{
tu_CmdEndRenderPass(commandBuffer);
}
projects / mesa.git / blobdiff