panfrost: Blit with the FullScreenJob descriptor (v9)

A vertex array is pre-allocated in order to interpolate varyings.

The GPU (tiler) only writes the VertexArrayDescriptor fields
"pointer", "vertex_packet_stride" and "vertex_attribute_stride" in
MallocVertexShaderJob mode. These fields can also be programmed with
IndexedVertexShaderJob and FullScreenJob modes. Update the XML spec
file comment accordingly.

Signed-off-by: Loïc Molinari <loic.molinari@collabora.com>
Reviewed-by: Ashley Smith <ashley.smith@collabora.com>
Acked-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/40124>

src/gallium/drivers/panfrost/pan_blit.c

@@ -25,8 +25,8 @@ panfrost_blitter_draw_rectangle(struct blitter_context *blitter,
    struct panfrost_context *pctx = pan_context(ctx);
    struct panfrost_screen *scr = pan_screen(ctx->screen);
 
-   /* Always fallback for now. */
-   goto fallback;
+   if (scr->dev.arch != 9 || depth != 0.0f || num_instances > 1)
+      goto fallback;
 
    /* Map viewport to the dest rect of the framebuffer. The tiler will then be
     * configured to use it as scissor box in order to clip fullscreen

src/gallium/drivers/panfrost/pan_context.h

@@ -435,6 +435,27 @@ struct panfrost_shader_binary {
    struct util_dynarray binary;
 };
 
+struct panfrost_run_fullscreen_attrib {
+   float x, y, z, w;
+};
+
+/* The tiler always allocates packets that can hold 64 vertices in RUN_IDVS
+ * malloc mode. For RUN_FULLSCREEN, the vertex array is preallocated but must
+ * match the tiler allocation strategy. */
+#define PAN_RUN_FULLSCREEN_NUM_VERTICES 64
+
+#define PAN_RUN_FULLSCREEN_ATTRIB_STRIDE \
+   sizeof(struct panfrost_run_fullscreen_attrib)
+
+/* A RUN_FULLSCREEN packet is made of a position and a texcoord attrib. */
+#define PAN_RUN_FULLSCREEN_PACKET_STRIDE \
+   (2 * sizeof(struct panfrost_run_fullscreen_attrib))
+
+#define PAN_RUN_FULLSCREEN_ARRAY_SIZE \
+   (PAN_RUN_FULLSCREEN_NUM_VERTICES * PAN_RUN_FULLSCREEN_PACKET_STRIDE)
+
+#define PAN_RUN_FULLSCREEN_ARRAY_ALIGN 64
+
 void
 panfrost_disk_cache_store(struct disk_cache *cache,
                           const struct panfrost_uncompiled_shader *uncompiled,
@@ -472,6 +493,10 @@ unsigned pan_assign_vertex_buffer(struct pan_vertex_buffer *buffers,
                                   unsigned *nr_bufs, unsigned vbi,
                                   unsigned divisor);
 
+struct pan_ptr panfrost_emit_fullscreen_vertex_array(struct panfrost_batch *batch,
+                                                     enum blitter_attrib_type type,
+                                                     const struct blitter_attrib *attrib);
+
 struct panfrost_zsa_state;
 struct panfrost_sampler_state;
 struct panfrost_sampler_view;

src/gallium/drivers/panfrost/pan_helpers.c

@@ -172,6 +172,41 @@ pan_assign_vertex_buffer(struct pan_vertex_buffer *buffers, unsigned *nr_bufs,
    return idx;
 }
 
+struct pan_ptr
+panfrost_emit_fullscreen_vertex_array(struct panfrost_batch *batch,
+                                      enum blitter_attrib_type type,
+                                      const struct blitter_attrib *attrib)
+{
+   struct pan_ptr array = { .cpu = NULL, .gpu = 0 };
+   struct panfrost_run_fullscreen_attrib *texcoords;
+
+   if (type != UTIL_BLITTER_ATTRIB_TEXCOORD_XY &&
+       type != UTIL_BLITTER_ATTRIB_TEXCOORD_XYZW)
+      return array;
+
+   array = pan_pool_alloc_aligned(&batch->pool.base,
+                                  PAN_RUN_FULLSCREEN_ARRAY_SIZE,
+                                  PAN_RUN_FULLSCREEN_ARRAY_ALIGN);
+   texcoords = (struct panfrost_run_fullscreen_attrib *)
+      ((uint8_t *)array.cpu + (PAN_RUN_FULLSCREEN_NUM_VERTICES *
+                               PAN_RUN_FULLSCREEN_ATTRIB_STRIDE));
+
+   texcoords[0].x = attrib->texcoord.x1;
+   texcoords[0].y = attrib->texcoord.y1;
+   texcoords[0].z = attrib->texcoord.z;
+   texcoords[0].w = attrib->texcoord.w;
+   texcoords[1].x = attrib->texcoord.x2;
+   texcoords[1].y = attrib->texcoord.y1;
+   texcoords[1].z = attrib->texcoord.z;
+   texcoords[1].w = attrib->texcoord.w;
+   texcoords[2].x = attrib->texcoord.x1;
+   texcoords[2].y = attrib->texcoord.y2;
+   texcoords[2].z = attrib->texcoord.z;
+   texcoords[2].w = attrib->texcoord.w;
+
+   return array;
+}
+
 /*
  * Helper to add a PIPE_CLEAR_* to batch->draws and batch->resolve together,
  * meaning that we draw to a given target. Adding to only one mask does not

src/gallium/drivers/panfrost/pan_jm.c

@@ -497,7 +497,8 @@ jm_emit_vertex_job(struct panfrost_batch *batch,
 
 static void
 jm_emit_tiler_draw(struct mali_draw_packed *out, struct panfrost_batch *batch,
-                   bool fs_required, enum mesa_prim prim)
+                   bool fs_required, enum mesa_prim prim,
+                   struct pan_ptr *vertex_array)
 {
    struct panfrost_context *ctx = batch->ctx;
    struct pipe_rasterizer_state *rast = &ctx->rasterizer->base;
@@ -559,6 +560,13 @@ jm_emit_tiler_draw(struct mali_draw_packed *out, struct panfrost_batch *batch,
       cfg.flags_0.aligned_line_ends = !rast->line_rectangular;
 
       cfg.vertex_array.packet = true;
+      if (vertex_array) {
+         cfg.vertex_array.pointer = vertex_array->gpu;
+         cfg.vertex_array.vertex_packet_stride =
+            PAN_RUN_FULLSCREEN_PACKET_STRIDE;
+         cfg.vertex_array.vertex_attribute_stride =
+            PAN_RUN_FULLSCREEN_ATTRIB_STRIDE;
+      }
 
       cfg.minimum_z = batch->minimum_z;
       cfg.maximum_z = batch->maximum_z;
@@ -804,7 +812,7 @@ jm_emit_malloc_vertex_job(struct panfrost_batch *batch,
    }
 
    jm_emit_tiler_draw(pan_section_ptr(job, MALLOC_VERTEX_JOB, DRAW), batch,
-                      fs_required, u_reduced_prim(info->mode));
+                      fs_required, u_reduced_prim(info->mode), NULL);
 
    pan_section_pack(job, MALLOC_VERTEX_JOB, POSITION, cfg) {
       jm_emit_shader_env(batch, &cfg, MESA_SHADER_VERTEX,
@@ -852,7 +860,8 @@ jm_emit_tiler_job(struct panfrost_batch *batch,
       ;
 #endif
 
-   jm_emit_tiler_draw(pan_section_ptr(job, TILER_JOB, DRAW), batch, true, prim);
+   jm_emit_tiler_draw(pan_section_ptr(job, TILER_JOB, DRAW), batch, true,
+                      prim, NULL);
 
    panfrost_emit_primitive_size(ctx, prim == MESA_PRIM_POINTS,
                                 batch->varyings.psiz, prim_size);
@@ -1025,7 +1034,44 @@ GENX(jm_launch_draw_fullscreen)(struct panfrost_batch *batch,
                                 enum blitter_attrib_type type,
                                 const struct blitter_attrib *attrib)
 {
-   UNREACHABLE("draw fullscreen not implemented for jm");
+#if PAN_ARCH == 9
+   PAN_TRACE_FUNC(PAN_TRACE_GL_JM);
+
+   struct pan_ptr job, dcd, vertex_array;
+
+   job = pan_pool_alloc_desc(&batch->pool.base, FULLSCREEN_JOB);
+   if (!job.cpu) {
+      mesa_loge("jm_launch_draw failed");
+      return;
+   }
+
+   dcd = pan_pool_alloc_desc(&batch->pool.base, DRAW);
+   if (!dcd.cpu) {
+      mesa_loge("jm_launch_draw failed");
+      return;
+   }
+
+   vertex_array = panfrost_emit_fullscreen_vertex_array(batch, type, attrib);
+   jm_emit_tiler_draw(dcd.cpu, batch, true, u_reduced_prim(MESA_PRIM_QUADS),
+                      &vertex_array);
+
+   pan_section_pack(job.cpu, FULLSCREEN_JOB, PRIMITIVE, cfg) {
+      cfg.scissor_array_enable = false;
+   }
+   pan_section_pack(job.cpu, FULLSCREEN_JOB, DCD, cfg) {
+      cfg.address = dcd.gpu;
+   }
+   pan_section_pack(job.cpu, FULLSCREEN_JOB, TILER, cfg) {
+      cfg.address = jm_emit_tiler_desc(batch);
+   }
+   memcpy(pan_section_ptr(job.cpu, FULLSCREEN_JOB, SCISSOR), &batch->scissor,
+          pan_size(SCISSOR));
+
+   pan_jc_add_job(&batch->jm.jobs.vtc_jc, MALI_JOB_TYPE_FULLSCREEN, false,
+                  false, 0, 0, &job, false);
+#else
+   UNREACHABLE("draw fullscreen not available for arch < 9");
+#endif
 }
 
 void

src/panfrost/genxml/v9.xml

@@ -43,6 +43,7 @@
     <value name="Fragment" value="9"/>
     <value name="Indexed vertex" value="10"/>
     <value name="Malloc vertex" value="11"/>
+    <value name="Fullscreen" value="12"/>
   </enum>
 
   <enum name="Shader stage">
@@ -1558,10 +1559,8 @@
   <struct name="Vertex Array" size="3">
     <field name="Packet" size="1" start="0:0" type="bool"/>
 
-    <!-- Written by hardware in packet mode -->
+    <!-- Written by hardware in MallocVertexShader job mode -->
     <field name="Pointer" size="58" start="0:6" type="address" modifier="shr(6)"/>
-
-    <!-- Written by hardware, leave zero -->
     <field name="Vertex packet stride" size="16" start="2:0" type="uint"/>
     <field name="Vertex attribute stride" size="16" start="2:16" type="uint"/>
   </struct>
@@ -1616,6 +1615,10 @@
     <field name="Vertex attribute stride" start="0:16" size="16" type="uint"/>
   </struct>
 
+  <struct name="DCD Pointer" size="2">
+    <field name="Address" size="64" start="0:0" type="address"/>
+  </struct>
+
   <struct name="Tiler Pointer" size="2">
     <field name="Address" size="64" start="0:0" type="address"/>
   </struct>
@@ -1643,6 +1646,14 @@
     <section name="Varying" offset="320" type="Shader Environment"/>
   </aggregate>
 
+  <aggregate name="Fullscreen Job" size="128" align="128">
+    <section name="Header" offset="0" type="Job Header"/>
+    <section name="Primitive" offset="32" type="Primitive"/>
+    <section name="DCD" offset="48" type="DCD Pointer"/>
+    <section name="Tiler" offset="56" type="Tiler Pointer"/>
+    <section name="Scissor" offset="104" type="Scissor"/>
+  </aggregate>
+
   <aggregate name="Tiler Job" size="256" align="128">
     <section name="Header" offset="0" type="Job Header"/>
     <section name="Primitive" offset="32" type="Primitive"/>

src/panfrost/lib/pan_jc.h

@@ -114,7 +114,8 @@ static inline bool
 job_uses_tiling(enum mali_job_type type)
 {
 #if PAN_ARCH >= 9
-   if (type == MALI_JOB_TYPE_MALLOC_VERTEX)
+   if (type == MALI_JOB_TYPE_MALLOC_VERTEX ||
+       type == MALI_JOB_TYPE_FULLSCREEN)
       return true;
 #endif