sema: r420c 2D fills/copies with pitch == row length cover all rows

victory_memset2d failed deterministically (bad ~= 12.5M of 12.58M) because
its pitched operations carry PITCH == LINE_LENGTH (cudaMallocPitch returned
pitch 4096 for width 4096) and the r402 2D decode only entered row mode
when a pitch was WIDER than the row. The flat path then ran ONE row
(LINE_LENGTH bytes) per op, so every cudaMemcpy2D readback and H2D upload
delivered only its first row and every cudaMemset2D-sized region beyond it
kept stale content - the 'deterministic readback loss' (45K of 12.58M
correct ~= the rows the single runs happened to cover).

LINE_COUNT is now authoritative for the flat case too: pitch == row length
(or no pitch method) means CONTIGUOUS rows, so the single run covers
rows*line bytes (element-exact when the line divides evenly, else byte
exact). A pitch wider than the row keeps the r402 row-wise mode; 1D ops are
unchanged. Both the fill (cudaMemset2D/fill) and the copy (cudaMemcpy2D)
decode are fixed.

1 files changed, 70 insertions, 22 deletions

diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index a465469..3b7c5e8 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -10000,18 +10000,45 @@ static void svc_decode_stream(struct nv_vgpu_device *vgpu, const u8 *buf,
 					 * r402: a 2D fill (cudaMemset2D):
 					 * LINE_COUNT rows of
 					 * LINE_LENGTH_IN bytes at
-					 * PITCH_OUT stride. Only a pitch
-					 * WIDER than the row selects row
-					 * mode; a flat fill carries
-					 * pitch == row length (or no
-					 * pitch method) and keeps the
-					 * historical single-run path.
+					 * PITCH_OUT stride. A pitch WIDER
+					 * than the row selects row mode; a
+					 * flat fill carries pitch == row
+					 * length (or no pitch method) and
+					 * keeps the historical single-run
+					 * path.
+					 *
+					 * r420c: LINE_COUNT is
+					 * authoritative for the flat case
+					 * too - cudaMemset2D with pitch ==
+					 * width (victory_memset2d,
+					 * pitch=4096/w=4096) is CONTIGUOUS
+					 * rows, so the single run must
+					 * cover rows*line bytes. The
+					 * historical path filled one row
+					 * and the readback lost the rest
+					 * (bad ~= 12.5M of 12.58M).
 					 */
-					if (st.have_pitch_out &&
-					    st.line_count > 1 &&
-					    st.pitch_out > st.line_length) {
-						m->rows = st.line_count;
-						m->dpitch = st.pitch_out;
+					if (st.line_count > 1) {
+						u32 fpitch = st.have_pitch_out ?
+							st.pitch_out :
+							st.line_length;
+
+						if (fpitch > st.line_length) {
+							m->rows =
+								st.line_count;
+							m->dpitch = fpitch;
+						} else if (st.line_length %
+							   m->elem) {
+							m->elem = 1;
+							m->count =
+								st.line_length *
+								st.line_count;
+						} else {
+							m->count =
+								(st.line_length /
+								 m->elem) *
+								st.line_count;
+						}
 					}
 					if (m->rows > 1 && svc_pitch_log < 32) {
 						svc_pitch_log++;
@@ -10053,20 +10080,41 @@ static void svc_decode_stream(struct nv_vgpu_device *vgpu, const u8 *buf,
 					 * LINE_COUNT rows of
 					 * LINE_LENGTH_IN bytes with
 					 * PITCH_IN/PITCH_OUT strides.
-					 * Row mode only when a pitch is
-					 * wider than the row (a flat 1D
-					 * copy keeps pitch == length).
+					 * Row mode when a pitch is
+					 * wider than the row.
+					 *
+					 * r420c: pitch == row length
+					 * (or no pitch method) makes the
+					 * rows CONTIGUOUS, so the single
+					 * flat run must cover
+					 * rows*line bytes - the
+					 * historical path copied one row
+					 * (victory_memset2d's pitched
+					 * D2H/H2D at pitch 4096 = width
+					 * 4096 lost all but the first
+					 * row).
 					 */
-					if (st.line_count > 1 &&
-					    st.have_pitch_out &&
-					    st.pitch_out > st.line_length) {
-						m->rows = st.line_count;
-						m->dpitch = st.pitch_out;
-						m->spitch = st.have_pitch_in ?
+					if (st.line_count > 1) {
+						u32 dpitch = st.have_pitch_out ?
+							st.pitch_out :
+							st.line_length;
+						u32 spitch = st.have_pitch_in ?
 							(st.pitch_in ?
 							 st.pitch_in :
-							 st.pitch_out) :
-							st.pitch_out;
+							 dpitch) :
+							dpitch;
+
+						if (dpitch == st.line_length &&
+						    spitch == st.line_length) {
+							m->bytes =
+								st.line_length *
+								st.line_count;
+						} else {
+							m->rows =
+								st.line_count;
+							m->dpitch = dpitch;
+							m->spitch = spitch;
+						}
 					}
 					n_ops++;
 					ops_seq++;