tu: Implement multiview clear/resolve interactions

Loads, stores, clears, and resolves now happen per-view. Since we only
support multiview with sysmem rendering, we only implement this for
sysmem clears and resolves.

There aren't any tests that mix multiview and MSAA, so no coverage of
the resolve path.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/5720>

diff --git a/src/freedreno/vulkan/tu_clear_blit.c b/src/freedreno/vulkan/tu_clear_blit.c
index e60f33b06df92f1d835a7d76e97018087d8fa961..fe39d6a40999e75f9ebddd682ed2ffe59070912f 100644 (file)
--- a/src/freedreno/vulkan/tu_clear_blit.c
+++ b/src/freedreno/vulkan/tu_clear_blit.c
@@ -439,6 +439,16 @@ r3d_common(struct tu_cmd_buffer *cmd, struct tu_cs *cs, bool blit, uint32_t num_
    tu_cs_emit_regs(cs, A6XX_PC_PRIMITIVE_CNTL_0());
    tu_cs_emit_regs(cs, A6XX_VFD_CONTROL_0());
 
+   /* Copy what the blob does here. This will emit an extra 0x3f
+    * CP_EVENT_WRITE when multiview is disabled. I'm not exactly sure what
+    * this is working around yet.
+    */
+   tu_cs_emit_pkt7(cs, CP_REG_WRITE, 3);
+   tu_cs_emit(cs, CP_REG_WRITE_0_TRACKER(UNK_EVENT_WRITE));
+   tu_cs_emit(cs, REG_A6XX_PC_MULTIVIEW_CNTL);
+   tu_cs_emit(cs, 0);
+   tu_cs_emit_regs(cs, A6XX_VFD_MULTIVIEW_CNTL());
+
    tu6_emit_vpc(cs, &vs, NULL, NULL, NULL, &fs, 0, false);
 
    /* REPL_MODE for varying with RECTLIST (2 vertices only) */
@@ -1665,11 +1675,18 @@ tu_CmdResolveImage(VkCommandBuffer commandBuffer,
    ops->teardown(cmd, cs);
 }
 
+#define for_each_layer(layer, layer_mask, layers) \
+   for (uint32_t layer = 0; \
+        layer < ((layer_mask) ? (util_logbase2(layer_mask) + 1) : layers); \
+        layer++) \
+      if (!layer_mask || (layer_mask & BIT(layer)))
+
 void
 tu_resolve_sysmem(struct tu_cmd_buffer *cmd,
                   struct tu_cs *cs,
                   struct tu_image_view *src,
                   struct tu_image_view *dst,
+                  uint32_t layer_mask,
                   uint32_t layers,
                   const VkRect2D *rect)
 {
@@ -1684,7 +1701,7 @@ tu_resolve_sysmem(struct tu_cmd_buffer *cmd,
               ROTATE_0, false, dst->ubwc_enabled);
    ops->coords(cs, &rect->offset, &rect->offset, &rect->extent);
 
-   for (uint32_t i = 0; i < layers; i++) {
+   for_each_layer(i, layer_mask, layers) {
       ops->src(cmd, cs, src, i, VK_FILTER_NEAREST);
       ops->dst(cs, dst, i);
       ops->run(cmd, cs);
@@ -1878,6 +1895,15 @@ tu_clear_sysmem_attachments(struct tu_cmd_buffer *cmd,
          layered_clear = true;
    }
 
+   /* a630 doesn't support multiview masks, which means that we can't use the
+    * normal multiview path without potentially recompiling a shader on-demand
+    * or using a more complicated variant that takes the mask as a const. Just
+    * use the layered path instead, since it shouldn't be much worse.
+    */
+   if (subpass->multiview_mask) {
+      layered_clear = true;
+   }
+
    r3d_common(cmd, cs, false, num_rts, layered_clear);
 
    tu_cs_emit_regs(cs,
@@ -1923,7 +1949,15 @@ tu_clear_sysmem_attachments(struct tu_cmd_buffer *cmd,
       tu_cs_emit_array(cs, clear_value[b], 4);
 
    for (uint32_t i = 0; i < rect_count; i++) {
-      for (uint32_t layer = 0; layer < rects[i].layerCount; layer++) {
+      /* This should be true because of this valid usage for
+       * vkCmdClearAttachments:
+       *
+       *    "If the render pass instance this is recorded in uses multiview,
+       *    then baseArrayLayer must be zero and layerCount must be one"
+       */
+      assert(!subpass->multiview_mask || rects[i].baseArrayLayer == 0);
+
+      for_each_layer(layer, subpass->multiview_mask, rects[i].layerCount) {
          r3d_coords_raw(cs, (float[]) {
             rects[i].rect.offset.x, rects[i].rect.offset.y,
             z_clear_val, uif(rects[i].baseArrayLayer + layer),
@@ -2150,6 +2184,7 @@ clear_sysmem_attachment(struct tu_cmd_buffer *cmd,
 {
    const struct tu_framebuffer *fb = cmd->state.framebuffer;
    const struct tu_image_view *iview = fb->attachments[a].attachment;
+   const uint32_t clear_views = cmd->state.pass->attachments[a].clear_views;
    const struct blit_ops *ops = &r2d_ops;
    if (cmd->state.pass->attachments[a].samples > 1)
       ops = &r3d_ops;
@@ -2158,7 +2193,7 @@ clear_sysmem_attachment(struct tu_cmd_buffer *cmd,
    ops->coords(cs, &info->renderArea.offset, NULL, &info->renderArea.extent);
    ops->clear_value(cs, format, &info->pClearValues[a]);
 
-   for (uint32_t i = 0; i < fb->layers; i++) {
+   for_each_layer(i, clear_views, fb->layers) {
       if (separate_stencil) {
          if (ops == &r3d_ops)
             r3d_dst_stencil(cs, iview, i);

diff --git a/src/freedreno/vulkan/tu_cmd_buffer.c b/src/freedreno/vulkan/tu_cmd_buffer.c
index fe1ddd83f73e966ea312be05fb1ab5ab4f839e0e..0a17add004c2e278cd4165d31139c33dc5aa7fba 100644 (file)
--- a/src/freedreno/vulkan/tu_cmd_buffer.c
+++ b/src/freedreno/vulkan/tu_cmd_buffer.c
@@ -340,7 +340,8 @@ tu6_emit_mrt(struct tu_cmd_buffer *cmd,
    tu_cs_emit_regs(cs,
                    A6XX_SP_SRGB_CNTL(.dword = subpass->srgb_cntl));
 
-   tu_cs_emit_regs(cs, A6XX_GRAS_MAX_LAYER_INDEX(fb->layers - 1));
+   unsigned layers = MAX2(fb->layers, util_logbase2(subpass->multiview_mask) + 1);
+   tu_cs_emit_regs(cs, A6XX_GRAS_MAX_LAYER_INDEX(layers - 1));
 }
 
 void
@@ -684,6 +685,7 @@ tu6_emit_tile_select(struct tu_cmd_buffer *cmd,
 static void
 tu6_emit_sysmem_resolve(struct tu_cmd_buffer *cmd,
                         struct tu_cs *cs,
+                        uint32_t layer_mask,
                         uint32_t a,
                         uint32_t gmem_a)
 {
@@ -691,7 +693,7 @@ tu6_emit_sysmem_resolve(struct tu_cmd_buffer *cmd,
    struct tu_image_view *dst = fb->attachments[a].attachment;
    struct tu_image_view *src = fb->attachments[gmem_a].attachment;
 
-   tu_resolve_sysmem(cmd, cs, src, dst, fb->layers, &cmd->state.render_area);
+   tu_resolve_sysmem(cmd, cs, src, dst, layer_mask, fb->layers, &cmd->state.render_area);
 }
 
 static void
@@ -731,7 +733,7 @@ tu6_emit_sysmem_resolves(struct tu_cmd_buffer *cmd,
          if (a == VK_ATTACHMENT_UNUSED)
             continue;
 
-         tu6_emit_sysmem_resolve(cmd, cs, a,
+         tu6_emit_sysmem_resolve(cmd, cs, subpass->multiview_mask, a,
                                  subpass->color_attachments[i].attachment);
       }
    }

diff --git a/src/freedreno/vulkan/tu_private.h b/src/freedreno/vulkan/tu_private.h
index 3d180fbecf69c1b99d3c7de616b64e8cb21b9434..584931dc534b39584f12c419508d79f56ac8a74e 100644 (file)
--- a/src/freedreno/vulkan/tu_private.h
+++ b/src/freedreno/vulkan/tu_private.h
@@ -1206,6 +1206,7 @@ tu_resolve_sysmem(struct tu_cmd_buffer *cmd,
                   struct tu_cs *cs,
                   struct tu_image_view *src,
                   struct tu_image_view *dst,
+                  uint32_t layer_mask,
                   uint32_t layers,
                   const VkRect2D *rect);