X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fgallium%2Fdrivers%2Ffreedreno%2Fa6xx%2Ffd6_blitter.c;h=4e4153ec718d59b0ea8e9c1e7e7cd0dca3c817bd;hb=d5c82c3c5f2f5e7a49630687173ff4f2debb3016;hp=97e015d19eb258c014eb233b8e9e4c0265206d44;hpb=0ac5acaeaaa651d850d77ebab094fa851a7ee06d;p=mesa.git

diff --git a/src/gallium/drivers/freedreno/a6xx/fd6_blitter.c b/src/gallium/drivers/freedreno/a6xx/fd6_blitter.c
index 97e015d19eb..4e4153ec718 100644
--- a/src/gallium/drivers/freedreno/a6xx/fd6_blitter.c
+++ b/src/gallium/drivers/freedreno/a6xx/fd6_blitter.c
@@ -26,14 +26,86 @@
  */
 
 #include "util/u_dump.h"
+#include "util/half_float.h"
 
 #include "freedreno_blitter.h"
 #include "freedreno_fence.h"
+#include "freedreno_log.h"
 #include "freedreno_resource.h"
 
 #include "fd6_blitter.h"
 #include "fd6_format.h"
 #include "fd6_emit.h"
+#include "fd6_resource.h"
+#include "fd6_pack.h"
+
+static inline enum a6xx_2d_ifmt
+fd6_ifmt(enum a6xx_format fmt)
+{
+	switch (fmt) {
+	case FMT6_A8_UNORM:
+	case FMT6_8_UNORM:
+	case FMT6_8_SNORM:
+	case FMT6_8_8_UNORM:
+	case FMT6_8_8_SNORM:
+	case FMT6_8_8_8_8_UNORM:
+	case FMT6_8_8_8_X8_UNORM:
+	case FMT6_8_8_8_8_SNORM:
+	case FMT6_4_4_4_4_UNORM:
+	case FMT6_5_5_5_1_UNORM:
+	case FMT6_5_6_5_UNORM:
+		return R2D_UNORM8;
+
+	case FMT6_32_UINT:
+	case FMT6_32_SINT:
+	case FMT6_32_32_UINT:
+	case FMT6_32_32_SINT:
+	case FMT6_32_32_32_32_UINT:
+	case FMT6_32_32_32_32_SINT:
+		return R2D_INT32;
+
+	case FMT6_16_UINT:
+	case FMT6_16_SINT:
+	case FMT6_16_16_UINT:
+	case FMT6_16_16_SINT:
+	case FMT6_16_16_16_16_UINT:
+	case FMT6_16_16_16_16_SINT:
+	case FMT6_10_10_10_2_UINT:
+		return R2D_INT16;
+
+	case FMT6_8_UINT:
+	case FMT6_8_SINT:
+	case FMT6_8_8_UINT:
+	case FMT6_8_8_SINT:
+	case FMT6_8_8_8_8_UINT:
+	case FMT6_8_8_8_8_SINT:
+	case FMT6_Z24_UNORM_S8_UINT:
+	case FMT6_Z24_UNORM_S8_UINT_AS_R8G8B8A8:
+		return R2D_INT8;
+
+	case FMT6_16_UNORM:
+	case FMT6_16_SNORM:
+	case FMT6_16_16_UNORM:
+	case FMT6_16_16_SNORM:
+	case FMT6_16_16_16_16_UNORM:
+	case FMT6_16_16_16_16_SNORM:
+	case FMT6_32_FLOAT:
+	case FMT6_32_32_FLOAT:
+	case FMT6_32_32_32_32_FLOAT:
+		return R2D_FLOAT32;
+
+	case FMT6_16_FLOAT:
+	case FMT6_16_16_FLOAT:
+	case FMT6_16_16_16_16_FLOAT:
+	case FMT6_11_11_10_FLOAT:
+	case FMT6_10_10_10_2_UNORM_DEST:
+		return R2D_FLOAT16;
+
+	default:
+		unreachable("bad format");
+		return 0;
+	}
+}
 
 /* Make sure none of the requested dimensions extend beyond the size of the
  * resource.  Not entirely sure why this happens, but sometimes it does, and
@@ -55,17 +127,33 @@ ok_dims(const struct pipe_resource *r, const struct pipe_box *b, int lvl)
 static bool
 ok_format(enum pipe_format pfmt)
 {
-	enum a6xx_color_fmt fmt = fd6_pipe2color(pfmt);
-	if (fmt == ~0)
-		return false;
+	enum a6xx_format fmt = fd6_pipe2color(pfmt);
+
+	if (util_format_is_compressed(pfmt))
+		return true;
+
+	switch (pfmt) {
+	case PIPE_FORMAT_Z24_UNORM_S8_UINT:
+	case PIPE_FORMAT_Z24X8_UNORM:
+	case PIPE_FORMAT_Z16_UNORM:
+	case PIPE_FORMAT_Z32_UNORM:
+	case PIPE_FORMAT_Z32_FLOAT:
+	case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
+	case PIPE_FORMAT_S8_UINT:
+		return true;
+	default:
+		break;
+	}
 
-	if (fd6_ifmt(fmt) == 0)
+	if (fmt == FMT6_NONE)
 		return false;
 
 	return true;
 }
 
+#define DEBUG_BLIT 0
 #define DEBUG_BLIT_FALLBACK 0
+
 #define fail_if(cond)													\
 	do {																\
 		if (cond) {														\
@@ -90,29 +178,12 @@ can_do_blit(const struct pipe_blit_info *info)
 	 */
 	fail_if(info->dst.box.depth != info->src.box.depth);
 
-	/* We can blit if both or neither formats are compressed formats... */
-	fail_if(util_format_is_compressed(info->src.format) !=
-			util_format_is_compressed(info->src.format));
-
 	/* Fail if unsupported format: */
 	fail_if(!ok_format(info->src.format));
 	fail_if(!ok_format(info->dst.format));
 
-	/* ... but only if they're the same compression format. */
-	fail_if(util_format_is_compressed(info->src.format) &&
-			info->src.format != info->dst.format);
-
-	/* hw ignores {SRC,DST}_INFO.COLOR_SWAP if {SRC,DST}_INFO.TILE_MODE
-	 * is set (not linear).  We can kind of get around that when tiling/
-	 * untiling by setting both src and dst COLOR_SWAP=WZYX, but that
-	 * means the formats must match:
-	 */
-	fail_if((fd_resource(info->dst.resource)->tile_mode ||
-			 fd_resource(info->src.resource)->tile_mode) &&
-			info->dst.format != info->src.format);
-
-	/* src box can be inverted, which we don't support.. dst box cannot: */
-	fail_if((info->src.box.width < 0) || (info->src.box.height < 0));
+	debug_assert(!util_format_is_compressed(info->src.format));
+	debug_assert(!util_format_is_compressed(info->dst.format));
 
 	fail_if(!ok_dims(info->src.resource, &info->src.box, info->src.level));
 
@@ -122,60 +193,113 @@ can_do_blit(const struct pipe_blit_info *info)
 	debug_assert(info->dst.box.height >= 0);
 	debug_assert(info->dst.box.depth >= 0);
 
-	fail_if(info->dst.resource->nr_samples + info->src.resource->nr_samples > 2);
+	fail_if(info->dst.resource->nr_samples > 1);
 
 	fail_if(info->window_rectangle_include);
 
-	fail_if(info->render_condition_enable);
+	const struct util_format_description *src_desc =
+		util_format_description(info->src.format);
+	const struct util_format_description *dst_desc =
+		util_format_description(info->dst.format);
+	const int common_channels = MIN2(src_desc->nr_channels, dst_desc->nr_channels);
+
+	if (info->mask & PIPE_MASK_RGBA) {
+		for (int i = 0; i < common_channels; i++) {
+			fail_if(memcmp(&src_desc->channel[i],
+						   &dst_desc->channel[i],
+						   sizeof(src_desc->channel[0])));
+		}
+	}
 
 	fail_if(info->alpha_blend);
 
-	fail_if(info->mask != util_format_get_mask(info->src.format));
-
-	fail_if(info->mask != util_format_get_mask(info->dst.format));
-
 	return true;
 }
 
 static void
-emit_setup(struct fd_ringbuffer *ring)
+emit_setup(struct fd_batch *batch)
 {
-	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
-	OUT_RING(ring, PC_CCU_INVALIDATE_COLOR);
-
-	OUT_PKT7(ring, CP_EVENT_WRITE, 1);
-	OUT_RING(ring, LRZ_FLUSH);
+	struct fd_ringbuffer *ring = batch->draw;
 
-	OUT_PKT7(ring, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
-	OUT_RING(ring, 0x0);
+	fd6_event_write(batch, ring, PC_CCU_FLUSH_COLOR_TS, true);
+	fd6_event_write(batch, ring, PC_CCU_FLUSH_DEPTH_TS, true);
+	fd6_event_write(batch, ring, PC_CCU_INVALIDATE_COLOR, false);
+	fd6_event_write(batch, ring, PC_CCU_INVALIDATE_DEPTH, false);
 
+	/* normal BLIT_OP_SCALE operation needs bypass RB_CCU_CNTL */
 	OUT_WFI5(ring);
-
 	OUT_PKT4(ring, REG_A6XX_RB_CCU_CNTL, 1);
-	OUT_RING(ring, 0x10000000);
+	OUT_RING(ring, fd6_context(batch->ctx)->magic.RB_CCU_CNTL_bypass);
 }
 
-static uint32_t
-blit_control(enum a6xx_color_fmt fmt)
+static void
+emit_blit_setup(struct fd_ringbuffer *ring,
+		enum pipe_format pfmt, bool scissor_enable, union pipe_color_union *color)
 {
-	unsigned blit_cntl = 0xf00000;
-	blit_cntl |= A6XX_RB_2D_BLIT_CNTL_COLOR_FORMAT(fmt);
-	blit_cntl |= A6XX_RB_2D_BLIT_CNTL_IFMT(fd6_ifmt(fmt));
-	return blit_cntl;
+	enum a6xx_format fmt = fd6_pipe2color(pfmt);
+	bool is_srgb = util_format_is_srgb(pfmt);
+	enum a6xx_2d_ifmt ifmt = fd6_ifmt(fmt);
+
+	OUT_PKT7(ring, CP_SET_MARKER, 1);
+	OUT_RING(ring, A6XX_CP_SET_MARKER_0_MODE(RM6_BLIT2DSCALE));
+
+	if (is_srgb) {
+		assert(ifmt == R2D_UNORM8);
+		ifmt = R2D_UNORM8_SRGB;
+	}
+
+	uint32_t blit_cntl = A6XX_RB_2D_BLIT_CNTL_MASK(0xf) |
+		A6XX_RB_2D_BLIT_CNTL_COLOR_FORMAT(fmt) |
+		A6XX_RB_2D_BLIT_CNTL_IFMT(ifmt) |
+		COND(color, A6XX_RB_2D_BLIT_CNTL_SOLID_COLOR) |
+		COND(scissor_enable, A6XX_RB_2D_BLIT_CNTL_SCISSOR);
+
+	OUT_PKT4(ring, REG_A6XX_RB_2D_BLIT_CNTL, 1);
+	OUT_RING(ring, blit_cntl);
+
+	OUT_PKT4(ring, REG_A6XX_GRAS_2D_BLIT_CNTL, 1);
+	OUT_RING(ring, blit_cntl);
+
+	if (fmt == FMT6_10_10_10_2_UNORM_DEST)
+		fmt = FMT6_16_16_16_16_FLOAT;
+
+	/* This register is probably badly named... it seems that it's
+	 * controlling the internal/accumulator format or something like
+	 * that. It's certainly not tied to only the src format.
+	 */
+	OUT_PKT4(ring, REG_A6XX_SP_2D_SRC_FORMAT, 1);
+	OUT_RING(ring, A6XX_SP_2D_SRC_FORMAT_COLOR_FORMAT(fmt) |
+			COND(util_format_is_pure_sint(pfmt),
+					A6XX_SP_2D_SRC_FORMAT_SINT) |
+			COND(util_format_is_pure_uint(pfmt),
+					A6XX_SP_2D_SRC_FORMAT_UINT) |
+			COND(util_format_is_snorm(pfmt),
+					A6XX_SP_2D_SRC_FORMAT_SINT |
+						A6XX_SP_2D_SRC_FORMAT_NORM) |
+			COND(util_format_is_unorm(pfmt),
+// TODO sometimes blob uses UINT+NORM but dEQP seems unhappy about that
+//						A6XX_SP_2D_SRC_FORMAT_UINT |
+					A6XX_SP_2D_SRC_FORMAT_NORM) |
+			COND(is_srgb, A6XX_SP_2D_SRC_FORMAT_SRGB) |
+			A6XX_SP_2D_SRC_FORMAT_MASK(0xf));
+
+	OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8C01, 1);
+	OUT_RING(ring, 0);
 }
 
 /* buffers need to be handled specially since x/width can exceed the bounds
  * supported by hw.. if necessary decompose into (potentially) two 2D blits
  */
 static void
-emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
+emit_blit_buffer(struct fd_context *ctx, struct fd_ringbuffer *ring,
+		const struct pipe_blit_info *info)
 {
 	const struct pipe_box *sbox = &info->src.box;
 	const struct pipe_box *dbox = &info->dst.box;
 	struct fd_resource *src, *dst;
 	unsigned sshift, dshift;
 
-	if (DEBUG_BLIT_FALLBACK) {
+	if (DEBUG_BLIT) {
 		fprintf(stderr, "buffer blit: ");
 		util_dump_blit_info(stderr, info);
 		fprintf(stderr, "\ndst resource: ");
@@ -188,8 +312,8 @@ emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 	src = fd_resource(info->src.resource);
 	dst = fd_resource(info->dst.resource);
 
-	debug_assert(src->cpp == 1);
-	debug_assert(dst->cpp == 1);
+	debug_assert(src->layout.cpp == 1);
+	debug_assert(dst->layout.cpp == 1);
 	debug_assert(info->src.resource->format == info->dst.resource->format);
 	debug_assert((sbox->y == 0) && (sbox->height == 1));
 	debug_assert((dbox->y == 0) && (dbox->height == 1));
@@ -220,15 +344,7 @@ emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 	sshift = sbox->x & 0x3f;
 	dshift = dbox->x & 0x3f;
 
-	OUT_PKT7(ring, CP_SET_MARKER, 1);
-	OUT_RING(ring, A2XX_CP_SET_MARKER_0_MODE(RM6_BLIT2DSCALE));
-
-	uint32_t blit_cntl = blit_control(RB6_R8_UNORM) | 0x20000000;
-	OUT_PKT4(ring, REG_A6XX_RB_2D_BLIT_CNTL, 1);
-	OUT_RING(ring, blit_cntl);
-
-	OUT_PKT4(ring, REG_A6XX_GRAS_2D_BLIT_CNTL, 1);
-	OUT_RING(ring, blit_cntl);
+	emit_blit_setup(ring, PIPE_FORMAT_R8_UNORM, false, NULL);
 
 	for (unsigned off = 0; off < sbox->width; off += (0x4000 - 0x40)) {
 		unsigned soff, doff, w, p;
@@ -245,10 +361,11 @@ emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 		/*
 		 * Emit source:
 		 */
-		OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_INFO, 13);
-		OUT_RING(ring, A6XX_SP_PS_2D_SRC_INFO_COLOR_FORMAT(RB6_R8_UNORM) |
+		OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_INFO, 10);
+		OUT_RING(ring, A6XX_SP_PS_2D_SRC_INFO_COLOR_FORMAT(FMT6_8_UNORM) |
 				A6XX_SP_PS_2D_SRC_INFO_TILE_MODE(TILE6_LINEAR) |
-				 A6XX_SP_PS_2D_SRC_INFO_COLOR_SWAP(WZYX) | 0x500000);
+				 A6XX_SP_PS_2D_SRC_INFO_COLOR_SWAP(WZYX) |
+				 0x500000);
 		OUT_RING(ring, A6XX_SP_PS_2D_SRC_SIZE_WIDTH(sshift + w) |
 				 A6XX_SP_PS_2D_SRC_SIZE_HEIGHT(1)); /* SP_PS_2D_SRC_SIZE */
 		OUT_RELOC(ring, src->bo, soff, 0, 0);    /* SP_PS_2D_SRC_LO/HI */
@@ -260,15 +377,11 @@ emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 		OUT_RING(ring, 0x00000000);
 		OUT_RING(ring, 0x00000000);
 
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-
 		/*
 		 * Emit destination:
 		 */
 		OUT_PKT4(ring, REG_A6XX_RB_2D_DST_INFO, 9);
-		OUT_RING(ring, A6XX_RB_2D_DST_INFO_COLOR_FORMAT(RB6_R8_UNORM) |
+		OUT_RING(ring, A6XX_RB_2D_DST_INFO_COLOR_FORMAT(FMT6_8_UNORM) |
 				 A6XX_RB_2D_DST_INFO_TILE_MODE(TILE6_LINEAR) |
 				 A6XX_RB_2D_DST_INFO_COLOR_SWAP(WZYX));
 		OUT_RELOC(ring, dst->bo, doff, 0, 0);    /* RB_2D_DST_LO/HI */
@@ -296,14 +409,8 @@ emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 		OUT_RING(ring, 0x3f);
 		OUT_WFI5(ring);
 
-		OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8C01, 1);
-		OUT_RING(ring, 0);
-
-		OUT_PKT4(ring, REG_A6XX_SP_2D_SRC_FORMAT, 1);
-		OUT_RING(ring, 0xf180);
-
 		OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1);
-		OUT_RING(ring, 0x01000000);
+		OUT_RING(ring, fd6_context(ctx)->magic.RB_UNKNOWN_8E04_blit);
 
 		OUT_PKT7(ring, CP_BLIT, 1);
 		OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_SCALE));
@@ -311,25 +418,110 @@ emit_blit_buffer(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 		OUT_WFI5(ring);
 
 		OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1);
-		OUT_RING(ring, 0);
+		OUT_RING(ring, 0);             /* RB_UNKNOWN_8E04 */
+	}
+}
+
+static void
+emit_blit_dst(struct fd_ringbuffer *ring, struct pipe_resource *prsc, enum pipe_format pfmt, unsigned level, unsigned layer)
+{
+	struct fd_resource *dst = fd_resource(prsc);
+	enum a6xx_format fmt = fd6_pipe2color(pfmt);
+	enum a6xx_tile_mode tile = fd_resource_tile_mode(prsc, level);
+	enum a3xx_color_swap swap = fd6_resource_swap(dst, pfmt);
+	uint32_t pitch = fd_resource_pitch(dst, level);
+	bool ubwc_enabled = fd_resource_ubwc_enabled(dst, level);
+	unsigned off = fd_resource_offset(dst, level, layer);
+
+	if (fmt == FMT6_Z24_UNORM_S8_UINT)
+		fmt = FMT6_Z24_UNORM_S8_UINT_AS_R8G8B8A8;
+
+	OUT_PKT4(ring, REG_A6XX_RB_2D_DST_INFO, 9);
+	OUT_RING(ring, A6XX_RB_2D_DST_INFO_COLOR_FORMAT(fmt) |
+			A6XX_RB_2D_DST_INFO_TILE_MODE(tile) |
+			A6XX_RB_2D_DST_INFO_COLOR_SWAP(swap) |
+			COND(util_format_is_srgb(pfmt), A6XX_RB_2D_DST_INFO_SRGB) |
+			COND(ubwc_enabled, A6XX_RB_2D_DST_INFO_FLAGS));
+	OUT_RELOC(ring, dst->bo, off, 0, 0);    /* RB_2D_DST_LO/HI */
+	OUT_RING(ring, A6XX_RB_2D_DST_SIZE_PITCH(pitch));
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+
+	if (ubwc_enabled) {
+		OUT_PKT4(ring, REG_A6XX_RB_2D_DST_FLAGS_LO, 6);
+		fd6_emit_flag_reference(ring, dst, level, layer);
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, 0x00000000);
 	}
 }
 
 static void
-emit_blit_texture(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
+emit_blit_src(struct fd_ringbuffer *ring, const struct pipe_blit_info *info, unsigned layer, unsigned nr_samples)
+{
+	struct fd_resource *src = fd_resource(info->src.resource);
+	enum a6xx_format sfmt = fd6_pipe2color(info->src.format);
+	enum a6xx_tile_mode	stile = fd_resource_tile_mode(info->src.resource, info->src.level);
+	enum a3xx_color_swap sswap = fd6_resource_swap(src, info->src.format);
+	uint32_t pitch = fd_resource_pitch(src, info->src.level);
+	bool subwc_enabled = fd_resource_ubwc_enabled(src, info->src.level);
+	unsigned soff = fd_resource_offset(src, info->src.level, layer);
+	uint32_t width = u_minify(src->base.width0, info->src.level) * nr_samples;
+	uint32_t height = u_minify(src->base.height0, info->src.level);
+	uint32_t filter = 0;
+
+	if (info->filter == PIPE_TEX_FILTER_LINEAR)
+		filter = A6XX_SP_PS_2D_SRC_INFO_FILTER;
+
+	enum a3xx_msaa_samples samples = fd_msaa_samples(src->base.nr_samples);
+
+	if (sfmt == FMT6_10_10_10_2_UNORM_DEST)
+		sfmt = FMT6_10_10_10_2_UNORM;
+
+	OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_INFO, 10);
+	OUT_RING(ring, A6XX_SP_PS_2D_SRC_INFO_COLOR_FORMAT(sfmt) |
+			A6XX_SP_PS_2D_SRC_INFO_TILE_MODE(stile) |
+			A6XX_SP_PS_2D_SRC_INFO_COLOR_SWAP(sswap) |
+			A6XX_SP_PS_2D_SRC_INFO_SAMPLES(samples) |
+			COND(samples > MSAA_ONE && (info->mask & PIPE_MASK_RGBA),
+					A6XX_SP_PS_2D_SRC_INFO_SAMPLES_AVERAGE) |
+			COND(subwc_enabled, A6XX_SP_PS_2D_SRC_INFO_FLAGS) |
+			COND(util_format_is_srgb(info->src.format), A6XX_SP_PS_2D_SRC_INFO_SRGB) |
+			0x500000 | filter);
+	OUT_RING(ring, A6XX_SP_PS_2D_SRC_SIZE_WIDTH(width) |
+			A6XX_SP_PS_2D_SRC_SIZE_HEIGHT(height)); /* SP_PS_2D_SRC_SIZE */
+	OUT_RELOC(ring, src->bo, soff, 0, 0);    /* SP_PS_2D_SRC_LO/HI */
+	OUT_RING(ring, A6XX_SP_PS_2D_SRC_PITCH_PITCH(pitch));
+
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+	OUT_RING(ring, 0x00000000);
+
+	if (subwc_enabled) {
+		OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_FLAGS_LO, 6);
+		fd6_emit_flag_reference(ring, src, info->src.level, layer);
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, 0x00000000);
+		OUT_RING(ring, 0x00000000);
+	}
+}
+
+static void
+emit_blit_texture(struct fd_context *ctx,
+		struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 {
 	const struct pipe_box *sbox = &info->src.box;
 	const struct pipe_box *dbox = &info->dst.box;
-	struct fd_resource *src, *dst;
-	struct fd_resource_slice *sslice, *dslice;
-	enum a6xx_color_fmt sfmt, dfmt;
-	enum a6xx_tile_mode stile, dtile;
-	enum a3xx_color_swap sswap, dswap;
-	unsigned spitch, dpitch;
-	unsigned sx1, sy1, sx2, sy2;
-	unsigned dx1, dy1, dx2, dy2;
-
-	if (DEBUG_BLIT_FALLBACK) {
+	struct fd_resource *dst;
+	int sx1, sy1, sx2, sy2;
+	int dx1, dy1, dx2, dy2;
+
+	if (DEBUG_BLIT) {
 		fprintf(stderr, "texture blit: ");
 		util_dump_blit_info(stderr, info);
 		fprintf(stderr, "\ndst resource: ");
@@ -339,167 +531,145 @@ emit_blit_texture(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 		fprintf(stderr, "\n");
 	}
 
-	src = fd_resource(info->src.resource);
 	dst = fd_resource(info->dst.resource);
 
-	sslice = fd_resource_slice(src, info->src.level);
-	dslice = fd_resource_slice(dst, info->dst.level);
+	uint32_t nr_samples = fd_resource_nr_samples(&dst->base);
 
-	sfmt = fd6_pipe2color(info->src.format);
-	dfmt = fd6_pipe2color(info->dst.format);
+	sx1 = sbox->x * nr_samples;
+	sy1 = sbox->y;
+	sx2 = (sbox->x + sbox->width) * nr_samples - 1;
+	sy2 = sbox->y + sbox->height - 1;
 
-	int blocksize = util_format_get_blocksize(info->src.format);
-	int blockwidth = util_format_get_blockwidth(info->src.format);
-	int blockheight = util_format_get_blockheight(info->src.format);
-	int nelements;
+	OUT_PKT4(ring, REG_A6XX_GRAS_2D_SRC_TL_X, 4);
+	OUT_RING(ring, A6XX_GRAS_2D_SRC_TL_X_X(sx1));
+	OUT_RING(ring, A6XX_GRAS_2D_SRC_BR_X_X(sx2));
+	OUT_RING(ring, A6XX_GRAS_2D_SRC_TL_Y_Y(sy1));
+	OUT_RING(ring, A6XX_GRAS_2D_SRC_BR_Y_Y(sy2));
 
-	stile = fd_resource_level_linear(info->src.resource, info->src.level) ?
-			TILE6_LINEAR : src->tile_mode;
-	dtile = fd_resource_level_linear(info->dst.resource, info->dst.level) ?
-			TILE6_LINEAR : dst->tile_mode;
+	dx1 = dbox->x * nr_samples;
+	dy1 = dbox->y;
+	dx2 = (dbox->x + dbox->width) * nr_samples - 1;
+	dy2 = dbox->y + dbox->height - 1;
 
-	sswap = fd6_pipe2swap(info->src.format);
-	dswap = fd6_pipe2swap(info->dst.format);
+	OUT_PKT4(ring, REG_A6XX_GRAS_2D_DST_TL, 2);
+	OUT_RING(ring, A6XX_GRAS_2D_DST_TL_X(dx1) | A6XX_GRAS_2D_DST_TL_Y(dy1));
+	OUT_RING(ring, A6XX_GRAS_2D_DST_BR_X(dx2) | A6XX_GRAS_2D_DST_BR_Y(dy2));
 
-	if (util_format_is_compressed(info->src.format)) {
-		debug_assert(info->src.format == info->dst.format);
-		sfmt = dfmt = RB6_R8_UNORM;
-		nelements = blocksize;
-	} else {
-		debug_assert(!util_format_is_compressed(info->dst.format));
-		nelements = 1;
+	if (info->scissor_enable) {
+		OUT_PKT4(ring, REG_A6XX_GRAS_RESOLVE_CNTL_1, 2);
+		OUT_RING(ring, A6XX_GRAS_RESOLVE_CNTL_1_X(info->scissor.minx) |
+				 A6XX_GRAS_RESOLVE_CNTL_1_Y(info->scissor.miny));
+		OUT_RING(ring, A6XX_GRAS_RESOLVE_CNTL_1_X(info->scissor.maxx - 1) |
+				 A6XX_GRAS_RESOLVE_CNTL_1_Y(info->scissor.maxy - 1));
 	}
 
-	spitch = DIV_ROUND_UP(sslice->pitch, blockwidth) * src->cpp;
-	dpitch = DIV_ROUND_UP(dslice->pitch, blockwidth) * dst->cpp;
-
-	sx1 = sbox->x / blockwidth * nelements;
-	sy1 = sbox->y / blockheight;
-	sx2 = DIV_ROUND_UP(sbox->x + sbox->width, blockwidth) * nelements - 1;
-	sy2 = DIV_ROUND_UP(sbox->y + sbox->height, blockheight) - 1;
+	emit_blit_setup(ring, info->dst.format, info->scissor_enable, NULL);
 
-	dx1 = dbox->x / blockwidth * nelements;
-	dy1 = dbox->y / blockheight;
-	dx2 = DIV_ROUND_UP(dbox->x + dbox->width, blockwidth) * nelements - 1;
-	dy2 = DIV_ROUND_UP(dbox->y + dbox->height, blockheight) - 1;
+	for (unsigned i = 0; i < info->dst.box.depth; i++) {
 
-	uint32_t width = DIV_ROUND_UP(u_minify(src->base.width0, info->src.level), blockwidth) * nelements;
-	uint32_t height = DIV_ROUND_UP(u_minify(src->base.height0, info->src.level), blockheight);
+		emit_blit_src(ring, info, sbox->z + i, nr_samples);
+		emit_blit_dst(ring, info->dst.resource, info->dst.format, info->dst.level, dbox->z + i);
 
-	/* if dtile, then dswap ignored by hw, and likewise if stile then sswap
-	 * ignored by hw.. but in this case we have already rejected the blit
-	 * if src and dst formats differ, so juse use WZYX for both src and
-	 * dst swap mode (so we don't change component order)
-	 */
-	if (stile || dtile) {
-		debug_assert(info->src.format == info->dst.format);
-		sswap = dswap = WZYX;
-	}
+		/*
+		 * Blit command:
+		 */
+		OUT_PKT7(ring, CP_EVENT_WRITE, 1);
+		OUT_RING(ring, 0x3f);
+		OUT_WFI5(ring);
 
-	OUT_PKT7(ring, CP_SET_MARKER, 1);
-	OUT_RING(ring, A2XX_CP_SET_MARKER_0_MODE(RM6_BLIT2DSCALE));
+		OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1);
+		OUT_RING(ring, fd6_context(ctx)->magic.RB_UNKNOWN_8E04_blit);
 
-	uint32_t blit_cntl = blit_control(dfmt);
+		OUT_PKT7(ring, CP_BLIT, 1);
+		OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_SCALE));
 
-	if (dtile != stile)
-		blit_cntl |= 0x20000000;
+		OUT_WFI5(ring);
 
-	if (info->scissor_enable) {
-		OUT_PKT4(ring, REG_A6XX_GRAS_RESOLVE_CNTL_1, 2);
-		OUT_RING(ring, A6XX_GRAS_RESOLVE_CNTL_1_X(info->scissor.minx) |
-				 A6XX_GRAS_RESOLVE_CNTL_1_Y(info->scissor.miny));
-		OUT_RING(ring, A6XX_GRAS_RESOLVE_CNTL_1_X(info->scissor.maxx - 1) |
-				 A6XX_GRAS_RESOLVE_CNTL_1_Y(info->scissor.maxy - 1));
-		blit_cntl |= A6XX_RB_2D_BLIT_CNTL_SCISSOR;
+		OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1);
+		OUT_RING(ring, 0);             /* RB_UNKNOWN_8E04 */
 	}
+}
 
-	OUT_PKT4(ring, REG_A6XX_RB_2D_BLIT_CNTL, 1);
-	OUT_RING(ring, blit_cntl);
+static void
+emit_clear_color(struct fd_ringbuffer *ring,
+		enum pipe_format pfmt, union pipe_color_union *color)
+{
+	switch (pfmt) {
+	case PIPE_FORMAT_Z24X8_UNORM:
+	case PIPE_FORMAT_Z24_UNORM_S8_UINT:
+	case PIPE_FORMAT_X24S8_UINT: {
+		uint32_t depth_unorm24 = color->f[0] * ((1u << 24) - 1);
+		uint8_t stencil = color->ui[1];
+		color->ui[0] = depth_unorm24 & 0xff;
+		color->ui[1] = (depth_unorm24 >> 8) & 0xff;
+		color->ui[2] = (depth_unorm24 >> 16) & 0xff;
+		color->ui[3] = stencil;
+		break;
+	}
+	default:
+		break;
+	}
 
-	OUT_PKT4(ring, REG_A6XX_GRAS_2D_BLIT_CNTL, 1);
-	OUT_RING(ring, blit_cntl);
+	OUT_PKT4(ring, REG_A6XX_RB_2D_SRC_SOLID_C0, 4);
+	switch (fd6_ifmt(fd6_pipe2color(pfmt))) {
+	case R2D_UNORM8:
+	case R2D_UNORM8_SRGB:
+		OUT_RING(ring, float_to_ubyte(color->f[0]));
+		OUT_RING(ring, float_to_ubyte(color->f[1]));
+		OUT_RING(ring, float_to_ubyte(color->f[2]));
+		OUT_RING(ring, float_to_ubyte(color->f[3]));
+		break;
+	case R2D_FLOAT16:
+		OUT_RING(ring, _mesa_float_to_half(color->f[0]));
+		OUT_RING(ring, _mesa_float_to_half(color->f[1]));
+		OUT_RING(ring, _mesa_float_to_half(color->f[2]));
+		OUT_RING(ring, _mesa_float_to_half(color->f[3]));
+		break;
+	case R2D_FLOAT32:
+	case R2D_INT32:
+	case R2D_INT16:
+	case R2D_INT8:
+	default:
+		OUT_RING(ring, color->ui[0]);
+		OUT_RING(ring, color->ui[1]);
+		OUT_RING(ring, color->ui[2]);
+		OUT_RING(ring, color->ui[3]);
+		break;
+	}
+}
 
-	for (unsigned i = 0; i < info->dst.box.depth; i++) {
-		unsigned soff = fd_resource_offset(src, info->src.level, sbox->z + i);
-		unsigned doff = fd_resource_offset(dst, info->dst.level, dbox->z + i);
+void
+fd6_clear_surface(struct fd_context *ctx,
+		struct fd_ringbuffer *ring, struct pipe_surface *psurf,
+		uint32_t width, uint32_t height, union pipe_color_union *color)
+{
+	if (DEBUG_BLIT) {
+		fprintf(stderr, "surface clear:\ndst resource: ");
+		util_dump_resource(stderr, psurf->texture);
+		fprintf(stderr, "\n");
+	}
 
-		/*
-		 * Emit source:
-		 */
-		uint32_t filter = 0;
-		if (info->filter == PIPE_TEX_FILTER_LINEAR)
-			filter = A6XX_SP_PS_2D_SRC_INFO_FILTER;
-
-		OUT_PKT4(ring, REG_A6XX_SP_PS_2D_SRC_INFO, 13);
-		OUT_RING(ring, A6XX_SP_PS_2D_SRC_INFO_COLOR_FORMAT(sfmt) |
-				A6XX_SP_PS_2D_SRC_INFO_TILE_MODE(stile) |
-				A6XX_SP_PS_2D_SRC_INFO_COLOR_SWAP(sswap) | 0x500000 | filter);
-		OUT_RING(ring, A6XX_SP_PS_2D_SRC_SIZE_WIDTH(width) |
-				 A6XX_SP_PS_2D_SRC_SIZE_HEIGHT(height)); /* SP_PS_2D_SRC_SIZE */
-		OUT_RELOC(ring, src->bo, soff, 0, 0);    /* SP_PS_2D_SRC_LO/HI */
-		OUT_RING(ring, A6XX_SP_PS_2D_SRC_PITCH_PITCH(spitch));
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
+	uint32_t nr_samples = fd_resource_nr_samples(psurf->texture);
+	OUT_PKT4(ring, REG_A6XX_GRAS_2D_DST_TL, 2);
+	OUT_RING(ring, A6XX_GRAS_2D_DST_TL_X(0) | A6XX_GRAS_2D_DST_TL_Y(0));
+	OUT_RING(ring, A6XX_GRAS_2D_DST_BR_X(width * nr_samples - 1) |
+			A6XX_GRAS_2D_DST_BR_Y(height - 1));
 
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
+	emit_clear_color(ring, psurf->format, color);
+	emit_blit_setup(ring, psurf->format, false, color);
 
-		/*
-		 * Emit destination:
-		 */
-		OUT_PKT4(ring, REG_A6XX_RB_2D_DST_INFO, 9);
-		OUT_RING(ring, A6XX_RB_2D_DST_INFO_COLOR_FORMAT(dfmt) |
-				 A6XX_RB_2D_DST_INFO_TILE_MODE(dtile) |
-				 A6XX_RB_2D_DST_INFO_COLOR_SWAP(dswap));
-		OUT_RELOC(ring, dst->bo, doff, 0, 0);    /* RB_2D_DST_LO/HI */
-		OUT_RING(ring, A6XX_RB_2D_DST_SIZE_PITCH(dpitch));
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
-		OUT_RING(ring, 0x00000000);
+	for (unsigned i = psurf->u.tex.first_layer; i <= psurf->u.tex.last_layer; i++) {
+		emit_blit_dst(ring, psurf->texture, psurf->format, psurf->u.tex.level, i);
 
 		/*
 		 * Blit command:
 		 */
-		OUT_PKT4(ring, REG_A6XX_GRAS_2D_SRC_TL_X, 4);
-		OUT_RING(ring, A6XX_GRAS_2D_SRC_TL_X_X(sx1));
-		OUT_RING(ring, A6XX_GRAS_2D_SRC_BR_X_X(sx2));
-		OUT_RING(ring, A6XX_GRAS_2D_SRC_TL_Y_Y(sy1));
-		OUT_RING(ring, A6XX_GRAS_2D_SRC_BR_Y_Y(sy2));
-
-		OUT_PKT4(ring, REG_A6XX_GRAS_2D_DST_TL, 2);
-		OUT_RING(ring, A6XX_GRAS_2D_DST_TL_X(dx1) | A6XX_GRAS_2D_DST_TL_Y(dy1));
-		OUT_RING(ring, A6XX_GRAS_2D_DST_BR_X(dx2) | A6XX_GRAS_2D_DST_BR_Y(dy2));
-
 		OUT_PKT7(ring, CP_EVENT_WRITE, 1);
 		OUT_RING(ring, 0x3f);
 		OUT_WFI5(ring);
 
-		OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8C01, 1);
-		OUT_RING(ring, 0);
-
-		OUT_PKT4(ring, REG_A6XX_SP_2D_SRC_FORMAT, 1);
-		OUT_RING(ring, A6XX_SP_2D_SRC_FORMAT_COLOR_FORMAT(sfmt) |
-				COND(util_format_is_pure_sint(info->src.format),
-						A6XX_SP_2D_SRC_FORMAT_SINT) |
-				COND(util_format_is_pure_uint(info->src.format),
-						A6XX_SP_2D_SRC_FORMAT_UINT) |
-				COND(util_format_is_snorm(info->src.format),
-						A6XX_SP_2D_SRC_FORMAT_SINT |
-						A6XX_SP_2D_SRC_FORMAT_NORM) |
-				COND(util_format_is_unorm(info->src.format),
-// TODO sometimes blob uses UINT+NORM but dEQP seems unhappy about that
-//						A6XX_SP_2D_SRC_FORMAT_UINT |
-						A6XX_SP_2D_SRC_FORMAT_NORM) |
-				0xf000);
-
 		OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1);
-		OUT_RING(ring, 0x01000000);
+		OUT_RING(ring, fd6_context(ctx)->magic.RB_UNKNOWN_8E04_blit);
 
 		OUT_PKT7(ring, CP_BLIT, 1);
 		OUT_RING(ring, CP_BLIT_0_OP(BLIT_OP_SCALE));
@@ -507,111 +677,228 @@ emit_blit_texture(struct fd_ringbuffer *ring, const struct pipe_blit_info *info)
 		OUT_WFI5(ring);
 
 		OUT_PKT4(ring, REG_A6XX_RB_UNKNOWN_8E04, 1);
-		OUT_RING(ring, 0);
+		OUT_RING(ring, 0);             /* RB_UNKNOWN_8E04 */
 	}
 }
 
-static void
-emit_blit(struct pipe_context *pctx, const struct pipe_blit_info *info)
+static bool
+handle_rgba_blit(struct fd_context *ctx, const struct pipe_blit_info *info)
 {
-	struct fd_context *ctx = fd_context(pctx);
 	struct fd_batch *batch;
 
-	fd_fence_ref(pctx->screen, &ctx->last_fence, NULL);
+	debug_assert(!(info->mask & PIPE_MASK_ZS));
+
+	if (!can_do_blit(info))
+		return false;
 
 	batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, true);
 
 	fd6_emit_restore(batch, batch->draw);
 	fd6_emit_lrz_flush(batch->draw);
 
-	mtx_lock(&ctx->screen->lock);
+	fd_screen_lock(ctx->screen);
 
-	fd_batch_resource_used(batch, fd_resource(info->src.resource), false);
-	fd_batch_resource_used(batch, fd_resource(info->dst.resource), true);
+	fd_batch_resource_read(batch, fd_resource(info->src.resource));
+	fd_batch_resource_write(batch, fd_resource(info->dst.resource));
+
+	fd_screen_unlock(ctx->screen);
+
+	/* Clearing last_fence must come after the batch dependency tracking
+	 * (resource_read()/resource_write()), as that can trigger a flush,
+	 * re-populating last_fence
+	 */
+	fd_fence_ref(&ctx->last_fence, NULL);
 
-	mtx_unlock(&ctx->screen->lock);
+	fd_batch_set_stage(batch, FD_STAGE_BLIT);
 
-	emit_setup(batch->draw);
+	fd_log_stream(batch, stream, util_dump_blit_info(stream, info));
+
+	emit_setup(batch);
 
 	if ((info->src.resource->target == PIPE_BUFFER) &&
 			(info->dst.resource->target == PIPE_BUFFER)) {
-		assert(fd_resource(info->src.resource)->tile_mode == TILE6_LINEAR);
-		assert(fd_resource(info->dst.resource)->tile_mode == TILE6_LINEAR);
-		emit_blit_buffer(batch->draw, info);
+		assert(fd_resource(info->src.resource)->layout.tile_mode == TILE6_LINEAR);
+		assert(fd_resource(info->dst.resource)->layout.tile_mode == TILE6_LINEAR);
+		fd_log(batch, "START BLIT (BUFFER)");
+		emit_blit_buffer(ctx, batch->draw, info);
+		fd_log(batch, "END BLIT (BUFFER)");
 	} else {
 		/* I don't *think* we need to handle blits between buffer <-> !buffer */
 		debug_assert(info->src.resource->target != PIPE_BUFFER);
 		debug_assert(info->dst.resource->target != PIPE_BUFFER);
-		emit_blit_texture(batch->draw, info);
+		fd_log(batch, "START BLIT (TEXTURE)");
+		emit_blit_texture(ctx, batch->draw, info);
+		fd_log(batch, "END BLIT (TEXTURE)");
 	}
 
-	fd6_event_write(batch, batch->draw, 0x1d, true);
-	fd6_event_write(batch, batch->draw, FACENESS_FLUSH, true);
+	fd6_event_write(batch, batch->draw, PC_CCU_FLUSH_COLOR_TS, true);
+	fd6_event_write(batch, batch->draw, PC_CCU_FLUSH_DEPTH_TS, true);
 	fd6_event_write(batch, batch->draw, CACHE_FLUSH_TS, true);
+	fd6_cache_inv(batch, batch->draw);
 
 	fd_resource(info->dst.resource)->valid = true;
 	batch->needs_flush = true;
 
-	fd_batch_flush(batch, false, false);
+	fd_batch_flush(batch);
 	fd_batch_reference(&batch, NULL);
+
+	return true;
 }
 
-static void
-fd6_blit(struct pipe_context *pctx, const struct pipe_blit_info *info)
+/**
+ * Re-written z/s blits can still fail for various reasons (for example MSAA).
+ * But we want to do the fallback blit with the re-written pipe_blit_info,
+ * in particular as u_blitter cannot blit stencil.  So handle the fallback
+ * ourself and never "fail".
+ */
+static bool
+do_rewritten_blit(struct fd_context *ctx, const struct pipe_blit_info *info)
 {
-	struct fd_context *ctx = fd_context(pctx);
+	bool success = handle_rgba_blit(ctx, info);
+	if (!success)
+		success = fd_blitter_blit(ctx, info);
+	debug_assert(success);  /* fallback should never fail! */
+	return success;
+}
 
-	if (!can_do_blit(info)) {
-		fd_blitter_pipe_begin(ctx, info->render_condition_enable, false, FD_STAGE_BLIT);
-		fd_blitter_blit(ctx, info);
-		fd_blitter_pipe_end(ctx);
-		return;
+/**
+ * Handle depth/stencil blits either via u_blitter and/or re-writing the
+ * blit into an equivilant format that we can handle
+ */
+static bool
+handle_zs_blit(struct fd_context *ctx, const struct pipe_blit_info *info)
+{
+	struct pipe_blit_info blit = *info;
+
+	if (DEBUG_BLIT) {
+		fprintf(stderr, "---- handle_zs_blit: ");
+		util_dump_blit_info(stderr, info);
+		fprintf(stderr, "\ndst resource: ");
+		util_dump_resource(stderr, info->dst.resource);
+		fprintf(stderr, "\nsrc resource: ");
+		util_dump_resource(stderr, info->src.resource);
+		fprintf(stderr, "\n");
 	}
 
-	emit_blit(pctx, info);
+	switch (info->dst.format) {
+	case PIPE_FORMAT_S8_UINT:
+		debug_assert(info->mask == PIPE_MASK_S);
+		blit.mask = PIPE_MASK_R;
+		blit.src.format = PIPE_FORMAT_R8_UINT;
+		blit.dst.format = PIPE_FORMAT_R8_UINT;
+		return do_rewritten_blit(ctx, &blit);
+
+	case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
+		if (info->mask & PIPE_MASK_Z) {
+			blit.mask = PIPE_MASK_R;
+			blit.src.format = PIPE_FORMAT_R32_FLOAT;
+			blit.dst.format = PIPE_FORMAT_R32_FLOAT;
+			do_rewritten_blit(ctx, &blit);
+		}
+
+		if (info->mask & PIPE_MASK_S) {
+			blit.mask = PIPE_MASK_R;
+			blit.src.format = PIPE_FORMAT_R8_UINT;
+			blit.dst.format = PIPE_FORMAT_R8_UINT;
+			blit.src.resource = &fd_resource(info->src.resource)->stencil->base;
+			blit.dst.resource = &fd_resource(info->dst.resource)->stencil->base;
+			do_rewritten_blit(ctx, &blit);
+		}
+
+		return true;
+
+	case PIPE_FORMAT_Z16_UNORM:
+		blit.mask = PIPE_MASK_R;
+		blit.src.format = PIPE_FORMAT_R16_UNORM;
+		blit.dst.format = PIPE_FORMAT_R16_UNORM;
+		return do_rewritten_blit(ctx, &blit);
+
+	case PIPE_FORMAT_Z32_UNORM:
+	case PIPE_FORMAT_Z32_FLOAT:
+		debug_assert(info->mask == PIPE_MASK_Z);
+		blit.mask = PIPE_MASK_R;
+		blit.src.format = PIPE_FORMAT_R32_UINT;
+		blit.dst.format = PIPE_FORMAT_R32_UINT;
+		return do_rewritten_blit(ctx, &blit);
+
+	case PIPE_FORMAT_Z24X8_UNORM:
+	case PIPE_FORMAT_Z24_UNORM_S8_UINT:
+		blit.mask = 0;
+		if (info->mask & PIPE_MASK_Z)
+			blit.mask |= PIPE_MASK_R | PIPE_MASK_G | PIPE_MASK_B;
+		if (info->mask & PIPE_MASK_S)
+			blit.mask |= PIPE_MASK_A;
+		blit.src.format = PIPE_FORMAT_Z24_UNORM_S8_UINT_AS_R8G8B8A8;
+		blit.dst.format = PIPE_FORMAT_Z24_UNORM_S8_UINT_AS_R8G8B8A8;
+		return fd_blitter_blit(ctx, &blit);
+
+	default:
+		return false;
+	}
 }
 
-static void
-fd6_resource_copy_region(struct pipe_context *pctx,
-		struct pipe_resource *dst,
-		unsigned dst_level,
-		unsigned dstx, unsigned dsty, unsigned dstz,
-		struct pipe_resource *src,
-		unsigned src_level,
-		const struct pipe_box *src_box)
+static bool
+handle_compressed_blit(struct fd_context *ctx, const struct pipe_blit_info *info)
 {
-	struct pipe_blit_info info;
-
-	debug_assert(src->format == dst->format);
-
-	memset(&info, 0, sizeof info);
-	info.dst.resource = dst;
-	info.dst.level = dst_level;
-	info.dst.box.x = dstx;
-	info.dst.box.y = dsty;
-	info.dst.box.z = dstz;
-	info.dst.box.width = src_box->width;
-	info.dst.box.height = src_box->height;
-	assert(info.dst.box.width >= 0);
-	assert(info.dst.box.height >= 0);
-	info.dst.box.depth = 1;
-	info.dst.format = dst->format;
-	info.src.resource = src;
-	info.src.level = src_level;
-	info.src.box = *src_box;
-	info.src.format = src->format;
-	info.mask = util_format_get_mask(src->format);
-	info.filter = PIPE_TEX_FILTER_NEAREST;
-	info.scissor_enable = 0;
-
-	if (!can_do_blit(&info)) {
-		fd_resource_copy_region(pctx,
-				dst, dst_level, dstx, dsty, dstz,
-				src, src_level, src_box);
-		return;
+	struct pipe_blit_info blit = *info;
+
+	if (DEBUG_BLIT) {
+		fprintf(stderr, "---- handle_compressed_blit: ");
+		util_dump_blit_info(stderr, info);
+		fprintf(stderr, "\ndst resource: ");
+		util_dump_resource(stderr, info->dst.resource);
+		fprintf(stderr, "\nsrc resource: ");
+		util_dump_resource(stderr, info->src.resource);
+		fprintf(stderr, "\n");
+	}
+
+	if (info->src.format != info->dst.format)
+		return fd_blitter_blit(ctx, info);
+
+	if (util_format_get_blocksize(info->src.format) == 8) {
+		blit.src.format = blit.dst.format = PIPE_FORMAT_R16G16B16A16_UINT;
+	} else {
+		debug_assert(util_format_get_blocksize(info->src.format) == 16);
+		blit.src.format = blit.dst.format = PIPE_FORMAT_R32G32B32A32_UINT;
 	}
 
-	emit_blit(pctx, &info);
+	int bw = util_format_get_blockwidth(info->src.format);
+	int bh = util_format_get_blockheight(info->src.format);
+
+	/* NOTE: x/y *must* be aligned to block boundary (ie. in
+	 * glCompressedTexSubImage2D()) but width/height may not
+	 * be:
+	 */
+
+	debug_assert((blit.src.box.x % bw) == 0);
+	debug_assert((blit.src.box.y % bh) == 0);
+
+	blit.src.box.x /= bw;
+	blit.src.box.y /= bh;
+	blit.src.box.width  = DIV_ROUND_UP(blit.src.box.width, bw);
+	blit.src.box.height = DIV_ROUND_UP(blit.src.box.height, bh);
+
+	debug_assert((blit.dst.box.x % bw) == 0);
+	debug_assert((blit.dst.box.y % bh) == 0);
+
+	blit.dst.box.x /= bw;
+	blit.dst.box.y /= bh;
+	blit.dst.box.width  = DIV_ROUND_UP(blit.dst.box.width, bw);
+	blit.dst.box.height = DIV_ROUND_UP(blit.dst.box.height, bh);
+
+	return do_rewritten_blit(ctx, &blit);
+}
+
+static bool
+fd6_blit(struct fd_context *ctx, const struct pipe_blit_info *info)
+{
+	if (info->mask & PIPE_MASK_ZS)
+		return handle_zs_blit(ctx, info);
+	if (util_format_is_compressed(info->src.format) ||
+			util_format_is_compressed(info->dst.format))
+		return handle_compressed_blit(ctx, info);
+
+	return handle_rgba_blit(ctx, info);
 }
 
 void
@@ -620,15 +907,23 @@ fd6_blitter_init(struct pipe_context *pctx)
 	if (fd_mesa_debug & FD_DBG_NOBLIT)
 		return;
 
-	pctx->resource_copy_region = fd6_resource_copy_region;
-	pctx->blit = fd6_blit;
+	fd_context(pctx)->blit = fd6_blit;
 }
 
 unsigned
 fd6_tile_mode(const struct pipe_resource *tmpl)
 {
+	/* if the mipmap level 0 is still too small to be tiled, then don't
+	 * bother pretending:
+	 */
+	if (fd_resource_level_linear(tmpl, 0))
+		return TILE6_LINEAR;
+
 	/* basically just has to be a format we can blit, so uploads/downloads
 	 * via linear staging buffer works:
 	 */
-	return TILE6_3;
+	if (ok_format(tmpl->format))
+		return TILE6_3;
+
+	return TILE6_LINEAR;
 }