diff --git a/src/gallium/drivers/panfrost/pan_fragment.c b/src/gallium/drivers/panfrost/pan_fragment.c
index 3c7d983a620..ec1914074dc 100644
--- a/src/gallium/drivers/panfrost/pan_fragment.c
+++ b/src/gallium/drivers/panfrost/pan_fragment.c
@@ -88,22 +88,24 @@ panfrost_fragment_job(struct panfrost_batch *batch, bool has_draws)
assert(batch->maxx > batch->minx);
assert(batch->maxy > batch->miny);
- struct mali_payload_fragment payload = {
- .min_tile_coord = MALI_COORDINATE_TO_TILE_MIN(batch->minx, batch->miny),
- .max_tile_coord = MALI_COORDINATE_TO_TILE_MAX(batch->maxx, batch->maxy),
- .framebuffer = framebuffer,
- };
-
struct panfrost_transfer transfer =
panfrost_pool_alloc_aligned(&batch->pool,
- MALI_JOB_HEADER_LENGTH + sizeof(payload),
- 64);
+ MALI_FRAGMENT_JOB_LENGTH, 64);
- pan_pack(transfer.cpu, JOB_HEADER, header) {
+ pan_section_pack(transfer.cpu, FRAGMENT_JOB, HEADER, header) {
header.type = MALI_JOB_TYPE_FRAGMENT;
header.index = 1;
}
- memcpy(transfer.cpu + MALI_JOB_HEADER_LENGTH, &payload, sizeof(payload));
+ pan_section_pack(transfer.cpu, FRAGMENT_JOB, PAYLOAD, payload) {
+ payload.bound_min_x = batch->minx >> MALI_TILE_SHIFT;
+ payload.bound_min_y = batch->miny >> MALI_TILE_SHIFT;
+
+ /* Batch max values are inclusive, we need to subtract 1. */
+ payload.bound_max_x = (batch->maxx - 1) >> MALI_TILE_SHIFT;
+ payload.bound_max_y = (batch->maxy - 1) >> MALI_TILE_SHIFT;
+ payload.framebuffer = framebuffer;
+ }
+
return transfer.gpu;
}
diff --git a/src/panfrost/include/panfrost-job.h b/src/panfrost/include/panfrost-job.h
index a0021508107..62a6e9edcd2 100644
--- a/src/panfrost/include/panfrost-job.h
+++ b/src/panfrost/include/panfrost-job.h
@@ -436,47 +436,4 @@ FIXED_16(float x, bool allow_negative)
return (int) (x * 256.0);
}
-/* From presentations, 16x16 tiles externally. Use shift for fast computation
- * of tile numbers. */
-
-#define MALI_TILE_SHIFT 4
-#define MALI_TILE_LENGTH (1 << MALI_TILE_SHIFT)
-
-/* Tile coordinates are stored as a compact u32, as only 12 bits are needed to
- * each component. Notice that this provides a theoretical upper bound of (1 <<
- * 12) = 4096 tiles in each direction, addressing a maximum framebuffer of size
- * 65536x65536. Multiplying that together, times another four given that Mali
- * framebuffers are 32-bit ARGB8888, means that this upper bound would take 16
- * gigabytes of RAM just to store the uncompressed framebuffer itself, let
- * alone rendering in real-time to such a buffer.
- *
- * Nice job, guys.*/
-
-/* From mali_kbase_10969_workaround.c */
-#define MALI_X_COORD_MASK 0x00000FFF
-#define MALI_Y_COORD_MASK 0x0FFF0000
-
-/* Extract parts of a tile coordinate */
-
-#define MALI_TILE_COORD_X(coord) ((coord) & MALI_X_COORD_MASK)
-#define MALI_TILE_COORD_Y(coord) (((coord) & MALI_Y_COORD_MASK) >> 16)
-
-/* Helpers to generate tile coordinates based on the boundary coordinates in
- * screen space. So, with the bounds (0, 0) to (128, 128) for the screen, these
- * functions would convert it to the bounding tiles (0, 0) to (7, 7).
- * Intentional "off-by-one"; finding the tile number is a form of fencepost
- * problem. */
-
-#define MALI_MAKE_TILE_COORDS(X, Y) ((X) | ((Y) << 16))
-#define MALI_BOUND_TO_TILE(B, bias) ((B - bias) >> MALI_TILE_SHIFT)
-#define MALI_COORDINATE_TO_TILE(W, H, bias) MALI_MAKE_TILE_COORDS(MALI_BOUND_TO_TILE(W, bias), MALI_BOUND_TO_TILE(H, bias))
-#define MALI_COORDINATE_TO_TILE_MIN(W, H) MALI_COORDINATE_TO_TILE(W, H, 0)
-#define MALI_COORDINATE_TO_TILE_MAX(W, H) MALI_COORDINATE_TO_TILE(W, H, 1)
-
-struct mali_payload_fragment {
- u32 min_tile_coord;
- u32 max_tile_coord;
- mali_ptr framebuffer;
-} __attribute__((packed));
-
#endif /* __PANFROST_JOB_H__ */
diff --git a/src/panfrost/lib/decode.c b/src/panfrost/lib/decode.c
index c36c4a95a75..4f483be3cec 100644
--- a/src/panfrost/lib/decode.c
+++ b/src/panfrost/lib/decode.c
@@ -1399,14 +1399,15 @@ pandecode_vertex_or_tiler_job_mdg(const struct MALI_JOB_HEADER *h,
return sizeof(*v);
}
-static int
+static void
pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
- mali_ptr payload, int job_no,
- bool is_bifrost, unsigned gpu_id)
+ mali_ptr job, int job_no,
+ bool is_bifrost, unsigned gpu_id)
{
- const struct mali_payload_fragment *PANDECODE_PTR_VAR(s, mem, payload);
+ struct mali_fragment_job_packed *PANDECODE_PTR_VAR(p, mem, job);
+ pan_section_unpack(p, FRAGMENT_JOB, PAYLOAD, s);
- bool is_mfbd = s->framebuffer & MALI_FBD_TAG_IS_MFBD;
+ bool is_mfbd = s.framebuffer & MALI_FBD_TAG_IS_MFBD;
if (!is_mfbd && is_bifrost)
pandecode_msg("XXX: Bifrost fragment must use MFBD\n");
@@ -1414,10 +1415,10 @@ pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
struct pandecode_fbd info;
if (is_mfbd)
- info = pandecode_mfbd_bfr(s->framebuffer & ~MALI_FBD_TAG_MASK, job_no,
+ info = pandecode_mfbd_bfr(s.framebuffer & ~MALI_FBD_TAG_MASK, job_no,
true, false, is_bifrost, gpu_id);
else
- info = pandecode_sfbd(s->framebuffer & ~MALI_FBD_TAG_MASK, job_no,
+ info = pandecode_sfbd(s.framebuffer & ~MALI_FBD_TAG_MASK, job_no,
true, gpu_id);
/* Compute the tag for the tagged pointer. This contains the type of
@@ -1434,35 +1435,19 @@ pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
expected_tag |= (MALI_POSITIVE(info.rt_count) << 2);
}
- if ((s->min_tile_coord | s->max_tile_coord) & ~(MALI_X_COORD_MASK | MALI_Y_COORD_MASK)) {
- pandecode_msg("XXX: unexpected tile coordinate bits\n");
- pandecode_prop("min_tile_coord = 0x%X\n", s->min_tile_coord);
- pandecode_prop("max_tile_coord = 0x%X\n", s->max_tile_coord);
- }
-
/* Extract tile coordinates */
- unsigned min_x = MALI_TILE_COORD_X(s->min_tile_coord) << MALI_TILE_SHIFT;
- unsigned min_y = MALI_TILE_COORD_Y(s->min_tile_coord) << MALI_TILE_SHIFT;
-
- unsigned max_x = (MALI_TILE_COORD_X(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
- unsigned max_y = (MALI_TILE_COORD_Y(s->max_tile_coord) + 1) << MALI_TILE_SHIFT;
-
- /* For the max, we also want the floored (rather than ceiled) version for checking */
-
- unsigned max_x_f = (MALI_TILE_COORD_X(s->max_tile_coord)) << MALI_TILE_SHIFT;
- unsigned max_y_f = (MALI_TILE_COORD_Y(s->max_tile_coord)) << MALI_TILE_SHIFT;
+ unsigned min_x = s.bound_min_x << MALI_TILE_SHIFT;
+ unsigned min_y = s.bound_min_y << MALI_TILE_SHIFT;
+ unsigned max_x = s.bound_max_x << MALI_TILE_SHIFT;
+ unsigned max_y = s.bound_max_y << MALI_TILE_SHIFT;
/* Validate the coordinates are well-ordered */
- if (min_x == max_x)
- pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
- else if (min_x > max_x)
+ if (min_x > max_x)
pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
- if (min_y == max_y)
- pandecode_msg("XXX: empty X coordinates (%u = %u)\n", min_x, max_x);
- else if (min_y > max_y)
+ if (min_y > max_y)
pandecode_msg("XXX: misordered X coordinates (%u > %u)\n", min_x, max_x);
/* Validate the coordinates fit inside the framebuffer. We use floor,
@@ -1470,24 +1455,24 @@ pandecode_fragment_job(const struct pandecode_mapped_memory *mem,
* coordinates for something like an 800x600 framebuffer will actually
* resolve to 800x608, which would otherwise trigger a Y-overflow */
- if ((min_x > info.width) || (max_x_f > info.width))
+ if (max_x + 1 > info.width)
pandecode_msg("XXX: tile coordinates overflow in X direction\n");
- if ((min_y > info.height) || (max_y_f > info.height))
+ if (max_y + 1 > info.height)
pandecode_msg("XXX: tile coordinates overflow in Y direction\n");
/* After validation, we print */
-
- pandecode_log("fragment (%u, %u) ... (%u, %u)\n\n", min_x, min_y, max_x, max_y);
+ DUMP_UNPACKED(FRAGMENT_JOB_PAYLOAD, s, "Fragment Job Payload:\n",
+ job + MALI_JOB_HEADER_LENGTH, job_no);
/* The FBD is a tagged pointer */
- unsigned tag = (s->framebuffer & MALI_FBD_TAG_MASK);
+ unsigned tag = (s.framebuffer & MALI_FBD_TAG_MASK);
if (tag != expected_tag)
pandecode_msg("XXX: expected FBD tag %X but got %X\n", expected_tag, tag);
- return sizeof(*s);
+ pandecode_log("\n");
}
static void
@@ -1554,7 +1539,7 @@ pandecode_jc(mali_ptr jc_gpu_va, bool bifrost, unsigned gpu_id, bool minimal)
break;
case MALI_JOB_TYPE_FRAGMENT:
- pandecode_fragment_job(mem, payload_ptr, job_no, bifrost, gpu_id);
+ pandecode_fragment_job(mem, jc_gpu_va, job_no, bifrost, gpu_id);
break;
default:
diff --git a/src/panfrost/lib/gen_pack.py b/src/panfrost/lib/gen_pack.py
index 2eb1936b785..cf212e9fb1c 100644
--- a/src/panfrost/lib/gen_pack.py
+++ b/src/panfrost/lib/gen_pack.py
@@ -163,6 +163,12 @@ __gen_unpack_padded(const uint8_t *restrict cl, uint32_t start, uint32_t end)
#define pan_section_print(fp, A, S, var, indent) \\
MALI_ ## A ## _SECTION_ ## S ## _print(fp, &(var), indent)
+/* From presentations, 16x16 tiles externally. Use shift for fast computation
+ * of tile numbers. */
+
+#define MALI_TILE_SHIFT 4
+#define MALI_TILE_LENGTH (1 << MALI_TILE_SHIFT)
+
"""
def to_alphanum(name):
diff --git a/src/panfrost/lib/midgard.xml b/src/panfrost/lib/midgard.xml
index 7154b859a42..d8c17de17b9 100644
--- a/src/panfrost/lib/midgard.xml
+++ b/src/panfrost/lib/midgard.xml
@@ -1060,6 +1060,22 @@
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+