openwrt_lede/target/linux/brcm2708/patches-4.4/0116-drm-vc4-Add-support-for-MSAA-rendering.patch

From c682c50a4564b43a43791dc03d0aa7c8848b2564 Mon Sep 17 00:00:00 2001
From: Eric Anholt <eric@anholt.net>
Date: Fri, 17 Jul 2015 13:15:50 -0700
Subject: [PATCH] drm/vc4: Add support for MSAA rendering.

For MSAA, you set a bit in the binner that halves the size of tiles in
each direction, so you can pack 4 samples per pixel in the tile
buffer.  During rendering, you can load and store raw tile buffer
contents (to save the per-sample MSAA contents), or you can load/store
resolved tile buffer contents (loads spam the pixel value to all 4
samples, and stores either average the 4 color samples, or store the
first sample for Z/S).

Signed-off-by: Eric Anholt <eric@anholt.net>
---
 drivers/gpu/drm/vc4/vc4_packet.h    |  23 ++-
 drivers/gpu/drm/vc4/vc4_render_cl.c | 274 ++++++++++++++++++++++++++++++------
 drivers/gpu/drm/vc4/vc4_validate.c  |   5 +-
 include/uapi/drm/vc4_drm.h          |  11 +-
 4 files changed, 258 insertions(+), 55 deletions(-)

--- a/drivers/gpu/drm/vc4/vc4_packet.h
+++ b/drivers/gpu/drm/vc4/vc4_packet.h
@@ -123,6 +123,11 @@ enum vc4_packet {
 #define VC4_PACKET_TILE_COORDINATES_SIZE				3
 #define VC4_PACKET_GEM_HANDLES_SIZE					9
 
+/* Number of multisamples supported. */
+#define VC4_MAX_SAMPLES							4
+/* Size of a full resolution color or Z tile buffer load/store. */
+#define VC4_TILE_BUFFER_SIZE			(64 * 64 * 4)
+
 /** @{
  * Bits used by packets like VC4_PACKET_STORE_TILE_BUFFER_GENERAL and
  * VC4_PACKET_TILE_RENDERING_MODE_CONFIG.
@@ -137,10 +142,20 @@ enum vc4_packet {
  * low bits of VC4_PACKET_STORE_FULL_RES_TILE_BUFFER and
  * VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER.
  */
-#define VC4_LOADSTORE_FULL_RES_EOF                     (1 << 3)
-#define VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL       (1 << 2)
-#define VC4_LOADSTORE_FULL_RES_DISABLE_ZS              (1 << 1)
-#define VC4_LOADSTORE_FULL_RES_DISABLE_COLOR           (1 << 0)
+#define VC4_LOADSTORE_FULL_RES_EOF                     BIT(3)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL       BIT(2)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_ZS              BIT(1)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_COLOR           BIT(0)
+
+/** @{
+ *
+ * low bits of VC4_PACKET_STORE_FULL_RES_TILE_BUFFER and
+ * VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER.
+ */
+#define VC4_LOADSTORE_FULL_RES_EOF                     BIT(3)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL       BIT(2)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_ZS              BIT(1)
+#define VC4_LOADSTORE_FULL_RES_DISABLE_COLOR           BIT(0)
 
 /** @{
  *
--- a/drivers/gpu/drm/vc4/vc4_render_cl.c
+++ b/drivers/gpu/drm/vc4/vc4_render_cl.c
@@ -37,9 +37,11 @@
 
 struct vc4_rcl_setup {
 	struct drm_gem_cma_object *color_read;
-	struct drm_gem_cma_object *color_ms_write;
+	struct drm_gem_cma_object *color_write;
 	struct drm_gem_cma_object *zs_read;
 	struct drm_gem_cma_object *zs_write;
+	struct drm_gem_cma_object *msaa_color_write;
+	struct drm_gem_cma_object *msaa_zs_write;
 
 	struct drm_gem_cma_object *rcl;
 	u32 next_offset;
@@ -82,6 +84,22 @@ static void vc4_store_before_load(struct
 }
 
 /*
+ * Calculates the physical address of the start of a tile in a RCL surface.
+ *
+ * Unlike the other load/store packets,
+ * VC4_PACKET_LOAD/STORE_FULL_RES_TILE_BUFFER don't look at the tile
+ * coordinates packet, and instead just store to the address given.
+ */
+static uint32_t vc4_full_res_offset(struct vc4_exec_info *exec,
+				    struct drm_gem_cma_object *bo,
+				    struct drm_vc4_submit_rcl_surface *surf,
+				    uint8_t x, uint8_t y)
+{
+	return bo->paddr + surf->offset + VC4_TILE_BUFFER_SIZE *
+		(DIV_ROUND_UP(exec->args->width, 32) * y + x);
+}
+
+/*
  * Emits a PACKET_TILE_COORDINATES if one isn't already pending.
  *
  * The tile coordinates packet triggers a pending load if there is one, are
@@ -108,22 +126,41 @@ static void emit_tile(struct vc4_exec_in
 	 * may be outstanding at a time.
 	 */
 	if (setup->color_read) {
-		rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
-		rcl_u16(setup, args->color_read.bits);
-		rcl_u32(setup,
-			setup->color_read->paddr + args->color_read.offset);
+		if (args->color_read.flags &
+		    VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+			rcl_u8(setup, VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER);
+			rcl_u32(setup,
+				vc4_full_res_offset(exec, setup->color_read,
+						    &args->color_read, x, y) |
+				VC4_LOADSTORE_FULL_RES_DISABLE_ZS);
+		} else {
+			rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
+			rcl_u16(setup, args->color_read.bits);
+			rcl_u32(setup, setup->color_read->paddr +
+				args->color_read.offset);
+		}
 	}
 
 	if (setup->zs_read) {
-		if (setup->color_read) {
-			/* Exec previous load. */
-			vc4_tile_coordinates(setup, x, y);
-			vc4_store_before_load(setup);
+		if (args->zs_read.flags &
+		    VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+			rcl_u8(setup, VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER);
+			rcl_u32(setup,
+				vc4_full_res_offset(exec, setup->zs_read,
+						    &args->zs_read, x, y) |
+				VC4_LOADSTORE_FULL_RES_DISABLE_COLOR);
+		} else {
+			if (setup->color_read) {
+				/* Exec previous load. */
+				vc4_tile_coordinates(setup, x, y);
+				vc4_store_before_load(setup);
+			}
+
+			rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
+			rcl_u16(setup, args->zs_read.bits);
+			rcl_u32(setup, setup->zs_read->paddr +
+				args->zs_read.offset);
 		}
-
-		rcl_u8(setup, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
-		rcl_u16(setup, args->zs_read.bits);
-		rcl_u32(setup, setup->zs_read->paddr + args->zs_read.offset);
 	}
 
 	/* Clipping depends on tile coordinates having been
@@ -144,20 +181,60 @@ static void emit_tile(struct vc4_exec_in
 				(y * exec->bin_tiles_x + x) * 32));
 	}
 
+	if (setup->msaa_color_write) {
+		bool last_tile_write = (!setup->msaa_zs_write &&
+					!setup->zs_write &&
+					!setup->color_write);
+		uint32_t bits = VC4_LOADSTORE_FULL_RES_DISABLE_ZS;
+
+		if (!last_tile_write)
+			bits |= VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL;
+		else if (last)
+			bits |= VC4_LOADSTORE_FULL_RES_EOF;
+		rcl_u8(setup, VC4_PACKET_STORE_FULL_RES_TILE_BUFFER);
+		rcl_u32(setup,
+			vc4_full_res_offset(exec, setup->msaa_color_write,
+					    &args->msaa_color_write, x, y) |
+			bits);
+	}
+
+	if (setup->msaa_zs_write) {
+		bool last_tile_write = (!setup->zs_write &&
+					!setup->color_write);
+		uint32_t bits = VC4_LOADSTORE_FULL_RES_DISABLE_COLOR;
+
+		if (setup->msaa_color_write)
+			vc4_tile_coordinates(setup, x, y);
+		if (!last_tile_write)
+			bits |= VC4_LOADSTORE_FULL_RES_DISABLE_CLEAR_ALL;
+		else if (last)
+			bits |= VC4_LOADSTORE_FULL_RES_EOF;
+		rcl_u8(setup, VC4_PACKET_STORE_FULL_RES_TILE_BUFFER);
+		rcl_u32(setup,
+			vc4_full_res_offset(exec, setup->msaa_zs_write,
+					    &args->msaa_zs_write, x, y) |
+			bits);
+	}
+
 	if (setup->zs_write) {
+		bool last_tile_write = !setup->color_write;
+
+		if (setup->msaa_color_write || setup->msaa_zs_write)
+			vc4_tile_coordinates(setup, x, y);
+
 		rcl_u8(setup, VC4_PACKET_STORE_TILE_BUFFER_GENERAL);
 		rcl_u16(setup, args->zs_write.bits |
-			(setup->color_ms_write ?
-			 VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR : 0));
+			(last_tile_write ?
+			 0 : VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR));
 		rcl_u32(setup,
 			(setup->zs_write->paddr + args->zs_write.offset) |
-			((last && !setup->color_ms_write) ?
+			((last && last_tile_write) ?
 			 VC4_LOADSTORE_TILE_BUFFER_EOF : 0));
 	}
 
-	if (setup->color_ms_write) {
-		if (setup->zs_write) {
-			/* Reset after previous store */
+	if (setup->color_write) {
+		if (setup->msaa_color_write || setup->msaa_zs_write ||
+		    setup->zs_write) {
 			vc4_tile_coordinates(setup, x, y);
 		}
 
@@ -192,14 +269,26 @@ static int vc4_create_rcl_bo(struct drm_
 	}
 
 	if (setup->color_read) {
-		loop_body_size += (VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE);
+		if (args->color_read.flags &
+		    VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+			loop_body_size += VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE;
+		} else {
+			loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
+		}
 	}
 	if (setup->zs_read) {
-		if (setup->color_read) {
-			loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE;
-			loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
+		if (args->zs_read.flags &
+		    VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+			loop_body_size += VC4_PACKET_LOAD_FULL_RES_TILE_BUFFER_SIZE;
+		} else {
+			if (setup->color_read &&
+			    !(args->color_read.flags &
+			      VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES)) {
+				loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE;
+				loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
+			}
+			loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
 		}
-		loop_body_size += VC4_PACKET_LOAD_TILE_BUFFER_GENERAL_SIZE;
 	}
 
 	if (has_bin) {
@@ -207,13 +296,23 @@ static int vc4_create_rcl_bo(struct drm_
 		loop_body_size += VC4_PACKET_BRANCH_TO_SUB_LIST_SIZE;
 	}
 
+	if (setup->msaa_color_write)
+		loop_body_size += VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE;
+	if (setup->msaa_zs_write)
+		loop_body_size += VC4_PACKET_STORE_FULL_RES_TILE_BUFFER_SIZE;
+
 	if (setup->zs_write)
 		loop_body_size += VC4_PACKET_STORE_TILE_BUFFER_GENERAL_SIZE;
-	if (setup->color_ms_write) {
-		if (setup->zs_write)
-			loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE;
+	if (setup->color_write)
 		loop_body_size += VC4_PACKET_STORE_MS_TILE_BUFFER_SIZE;
-	}
+
+	/* We need a VC4_PACKET_TILE_COORDINATES in between each store. */
+	loop_body_size += VC4_PACKET_TILE_COORDINATES_SIZE *
+		((setup->msaa_color_write != NULL) +
+		 (setup->msaa_zs_write != NULL) +
+		 (setup->color_write != NULL) +
+		 (setup->zs_write != NULL) - 1);
+
 	size += xtiles * ytiles * loop_body_size;
 
 	setup->rcl = &vc4_bo_create(dev, size, true)->base;
@@ -224,13 +323,12 @@ static int vc4_create_rcl_bo(struct drm_
 
 	rcl_u8(setup, VC4_PACKET_TILE_RENDERING_MODE_CONFIG);
 	rcl_u32(setup,
-		(setup->color_ms_write ?
-		 (setup->color_ms_write->paddr +
-		  args->color_ms_write.offset) :
+		(setup->color_write ? (setup->color_write->paddr +
+				       args->color_write.offset) :
 		 0));
 	rcl_u16(setup, args->width);
 	rcl_u16(setup, args->height);
-	rcl_u16(setup, args->color_ms_write.bits);
+	rcl_u16(setup, args->color_write.bits);
 
 	/* The tile buffer gets cleared when the previous tile is stored.  If
 	 * the clear values changed between frames, then the tile buffer has
@@ -267,6 +365,56 @@ static int vc4_create_rcl_bo(struct drm_
 	return 0;
 }
 
+static int vc4_full_res_bounds_check(struct vc4_exec_info *exec,
+				     struct drm_gem_cma_object *obj,
+				     struct drm_vc4_submit_rcl_surface *surf)
+{
+	struct drm_vc4_submit_cl *args = exec->args;
+	u32 render_tiles_stride = DIV_ROUND_UP(exec->args->width, 32);
+
+	if (surf->offset > obj->base.size) {
+		DRM_ERROR("surface offset %d > BO size %zd\n",
+			  surf->offset, obj->base.size);
+		return -EINVAL;
+	}
+
+	if ((obj->base.size - surf->offset) / VC4_TILE_BUFFER_SIZE <
+	    render_tiles_stride * args->max_y_tile + args->max_x_tile) {
+		DRM_ERROR("MSAA tile %d, %d out of bounds "
+			  "(bo size %zd, offset %d).\n",
+			  args->max_x_tile, args->max_y_tile,
+			  obj->base.size,
+			  surf->offset);
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
+static int vc4_rcl_msaa_surface_setup(struct vc4_exec_info *exec,
+				      struct drm_gem_cma_object **obj,
+				      struct drm_vc4_submit_rcl_surface *surf)
+{
+	if (surf->flags != 0 || surf->bits != 0) {
+		DRM_ERROR("MSAA surface had nonzero flags/bits\n");
+		return -EINVAL;
+	}
+
+	if (surf->hindex == ~0)
+		return 0;
+
+	*obj = vc4_use_bo(exec, surf->hindex);
+	if (!*obj)
+		return -EINVAL;
+
+	if (surf->offset & 0xf) {
+		DRM_ERROR("MSAA write must be 16b aligned.\n");
+		return -EINVAL;
+	}
+
+	return vc4_full_res_bounds_check(exec, *obj, surf);
+}
+
 static int vc4_rcl_surface_setup(struct vc4_exec_info *exec,
 				 struct drm_gem_cma_object **obj,
 				 struct drm_vc4_submit_rcl_surface *surf)
@@ -278,9 +426,10 @@ static int vc4_rcl_surface_setup(struct
 	uint8_t format = VC4_GET_FIELD(surf->bits,
 				       VC4_LOADSTORE_TILE_BUFFER_FORMAT);
 	int cpp;
+	int ret;
 
-	if (surf->pad != 0) {
-		DRM_ERROR("Padding unset\n");
+	if (surf->flags & ~VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+		DRM_ERROR("Extra flags set\n");
 		return -EINVAL;
 	}
 
@@ -290,6 +439,25 @@ static int vc4_rcl_surface_setup(struct
 	if (!vc4_use_bo(exec, surf->hindex, VC4_MODE_RENDER, obj))
 		return -EINVAL;
 
+	if (surf->flags & VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES) {
+		if (surf == &exec->args->zs_write) {
+			DRM_ERROR("general zs write may not be a full-res.\n");
+			return -EINVAL;
+		}
+
+		if (surf->bits != 0) {
+			DRM_ERROR("load/store general bits set with "
+				  "full res load/store.\n");
+			return -EINVAL;
+		}
+
+		ret = vc4_full_res_bounds_check(exec, *obj, surf);
+		if (!ret)
+			return ret;
+
+		return 0;
+	}
+
 	if (surf->bits & ~(VC4_LOADSTORE_TILE_BUFFER_TILING_MASK |
 			   VC4_LOADSTORE_TILE_BUFFER_BUFFER_MASK |
 			   VC4_LOADSTORE_TILE_BUFFER_FORMAT_MASK)) {
@@ -341,9 +509,10 @@ static int vc4_rcl_surface_setup(struct
 }
 
 static int
-vc4_rcl_ms_surface_setup(struct vc4_exec_info *exec,
-			 struct drm_gem_cma_object **obj,
-			 struct drm_vc4_submit_rcl_surface *surf)
+vc4_rcl_render_config_surface_setup(struct vc4_exec_info *exec,
+				    struct vc4_rcl_setup *setup,
+				    struct drm_gem_cma_object **obj,
+				    struct drm_vc4_submit_rcl_surface *surf)
 {
 	uint8_t tiling = VC4_GET_FIELD(surf->bits,
 				       VC4_RENDER_CONFIG_MEMORY_FORMAT);
@@ -351,13 +520,15 @@ vc4_rcl_ms_surface_setup(struct vc4_exec
 				       VC4_RENDER_CONFIG_FORMAT);
 	int cpp;
 
-	if (surf->pad != 0) {
-		DRM_ERROR("Padding unset\n");
+	if (surf->flags != 0) {
+		DRM_ERROR("No flags supported on render config.\n");
 		return -EINVAL;
 	}
 
 	if (surf->bits & ~(VC4_RENDER_CONFIG_MEMORY_FORMAT_MASK |
-			   VC4_RENDER_CONFIG_FORMAT_MASK)) {
+			   VC4_RENDER_CONFIG_FORMAT_MASK |
+			   VC4_RENDER_CONFIG_MS_MODE_4X |
+			   VC4_RENDER_CONFIG_DECIMATE_MODE_4X)) {
 		DRM_ERROR("Unknown bits in render config: 0x%04x\n",
 			  surf->bits);
 		return -EINVAL;
@@ -413,18 +584,20 @@ int vc4_get_rcl(struct drm_device *dev,
 	if (has_bin &&
 	    (args->max_x_tile > exec->bin_tiles_x ||
 	     args->max_y_tile > exec->bin_tiles_y)) {
-		DRM_ERROR("Render tiles (%d,%d) outside of bin config (%d,%d)\n",
+		DRM_ERROR("Render tiles (%d,%d) outside of bin config "
+			  "(%d,%d)\n",
 			  args->max_x_tile, args->max_y_tile,
 			  exec->bin_tiles_x, exec->bin_tiles_y);
 		return -EINVAL;
 	}
 
-	ret = vc4_rcl_surface_setup(exec, &setup.color_read, &args->color_read);
+	ret = vc4_rcl_render_config_surface_setup(exec, &setup,
+						  &setup.color_write,
+						  &args->color_write);
 	if (ret)
 		return ret;
 
-	ret = vc4_rcl_ms_surface_setup(exec, &setup.color_ms_write,
-				       &args->color_ms_write);
+	ret = vc4_rcl_surface_setup(exec, &setup.color_read, &args->color_read);
 	if (ret)
 		return ret;
 
@@ -436,10 +609,21 @@ int vc4_get_rcl(struct drm_device *dev,
 	if (ret)
 		return ret;
 
+	ret = vc4_rcl_msaa_surface_setup(exec, &setup.msaa_color_write,
+					 &args->msaa_color_write);
+	if (ret)
+		return ret;
+
+	ret = vc4_rcl_msaa_surface_setup(exec, &setup.msaa_zs_write,
+					 &args->msaa_zs_write);
+	if (ret)
+		return ret;
+
 	/* We shouldn't even have the job submitted to us if there's no
 	 * surface to write out.
 	 */
-	if (!setup.color_ms_write && !setup.zs_write) {
+	if (!setup.color_write && !setup.zs_write &&
+	    !setup.msaa_color_write && !setup.msaa_zs_write) {
 		DRM_ERROR("RCL requires color or Z/S write\n");
 		return -EINVAL;
 	}
--- a/drivers/gpu/drm/vc4/vc4_validate.c
+++ b/drivers/gpu/drm/vc4/vc4_validate.c
@@ -400,9 +400,8 @@ validate_tile_binning_config(VALIDATE_AR
 	}
 
 	if (flags & (VC4_BIN_CONFIG_DB_NON_MS |
-		     VC4_BIN_CONFIG_TILE_BUFFER_64BIT |
-		     VC4_BIN_CONFIG_MS_MODE_4X)) {
-		DRM_ERROR("unsupported bining config flags 0x%02x\n", flags);
+		     VC4_BIN_CONFIG_TILE_BUFFER_64BIT)) {
+		DRM_ERROR("unsupported binning config flags 0x%02x\n", flags);
 		return -EINVAL;
 	}
 
--- a/include/uapi/drm/vc4_drm.h
+++ b/include/uapi/drm/vc4_drm.h
@@ -46,10 +46,13 @@ struct drm_vc4_submit_rcl_surface {
 	uint32_t hindex; /* Handle index, or ~0 if not present. */
 	uint32_t offset; /* Offset to start of buffer. */
 	/*
-         * Bits for either render config (color_ms_write) or load/store packet.
+         * Bits for either render config (color_write) or load/store packet.
+         * Bits should all be 0 for MSAA load/stores.
 	 */
 	uint16_t bits;
-	uint16_t pad;
+
+#define VC4_SUBMIT_RCL_SURFACE_READ_IS_FULL_RES		(1 << 0)
+	uint16_t flags;
 };
 
 /**
@@ -128,9 +131,11 @@ struct drm_vc4_submit_cl {
 	uint8_t max_x_tile;
 	uint8_t max_y_tile;
 	struct drm_vc4_submit_rcl_surface color_read;
-	struct drm_vc4_submit_rcl_surface color_ms_write;
+	struct drm_vc4_submit_rcl_surface color_write;
 	struct drm_vc4_submit_rcl_surface zs_read;
 	struct drm_vc4_submit_rcl_surface zs_write;
+	struct drm_vc4_submit_rcl_surface msaa_color_write;
+	struct drm_vc4_submit_rcl_surface msaa_zs_write;
 	uint32_t clear_color[2];
 	uint32_t clear_z;
 	uint8_t clear_s;