vgsp-gl: GS SPH placement/maps/context A/B knobs

3 files changed, 77 insertions, 3 deletions

diff --git a/kernel/gsp_phys_booter.c b/kernel/gsp_phys_booter.c
index 0fc2ee7..96fdc46 100644
--- a/kernel/gsp_phys_booter.c
+++ b/kernel/gsp_phys_booter.c
@@ -18713,6 +18713,42 @@ bool gsp_phys_rmcall_gr_raster_pre_value(void)
 {
 	return READ_ONCE(gsp_phys_rmcall_gr_raster_pre);
 }
+
+/*
+ * SESSION 155 CONT 15: where the synthesized GS SPH lives inside the
+ * program-family page. 0 = family+0x780 (below the window's own VTG bind
+ * at family+0x900), 1 = family+0xe80 (above it; the DS's PROGRAM_REGION
+ * snapshot reads the bind VA, so a region starting there does not cover
+ * 0x780). Default 0 = legacy.
+ */
+static u32 gsp_phys_rmcall_gr_sph_gs_addr;
+module_param_named(sema_gl_sph_gs_addr, gsp_phys_rmcall_gr_sph_gs_addr,
+		   uint, 0644);
+MODULE_PARM_DESC(sema_gl_sph_gs_addr,
+	"Synth GS SPH offset in the family page: 0=+0x780 (legacy), 1=+0xe80 (above the VTG bind)");
+
+u32 gsp_phys_rmcall_gr_sph_gs_addr_value(void)
+{
+	return READ_ONCE(gsp_phys_rmcall_gr_sph_gs_addr);
+}
+
+/*
+ * SESSION 155 CONT 15: GS context A/B words emitted with the synth GS bind.
+ * bit0: SET_MESH_CONTROL (0x114c) <- 0
+ * bit1: SET_GS_MODE     (0x0208) <- 0 (TYPE_ANY)
+ * The client's per-frame push carries neither; if the init push left them
+ * set, the GS macro/mesh path may treat the stage specially. Default 0.
+ */
+static u32 gsp_phys_rmcall_gr_sph_gs_ctx;
+module_param_named(sema_gl_sph_gs_ctx, gsp_phys_rmcall_gr_sph_gs_ctx,
+		   uint, 0644);
+MODULE_PARM_DESC(sema_gl_sph_gs_ctx,
+	"Extra GS context with the synth bind: bit0=SET_MESH_CONTROL(0x114c)<-0, bit1=SET_GS_MODE(0x0208)<-0 (default 0)");
+
+u32 gsp_phys_rmcall_gr_sph_gs_ctx_value(void)
+{
+	return READ_ONCE(gsp_phys_rmcall_gr_sph_gs_ctx);
+}
 static bool gsp_phys_gr_mme_pairs_emitted;	/* launch emitted a pair run */
 static u32 gsp_phys_gr_mme_pairs_gen;		/* gen whose pairs retired OK */
 static u32 gsp_phys_gr_mme_launch_gen;		/* last built launch's gen */
@@ -29776,6 +29812,8 @@ gr_page_unmapped:
 				READ_ONCE(gsp_phys_rmcall_gr_sph_gs_mode);
 			bool emit_raster =
 				READ_ONCE(gsp_phys_rmcall_gr_raster_pre);
+			u32 gs_addr = READ_ONCE(gsp_phys_rmcall_gr_sph_gs_addr);
+			u32 gs_ctx = READ_ONCE(gsp_phys_rmcall_gr_sph_gs_ctx);
 
 			/*
 			 * SESSION 155 CONT 9: the prelude is prepended ahead of
@@ -30004,7 +30042,9 @@ gr_page_unmapped:
 			    (emit_j2 ? (bind_binding ? 8u : 7u) : 0u) +
 			    (emit_j4 ? (bind_binding ? 8u : 7u) : 0u) +
 			    (emit_j5 ? (bind_binding ? 8u : 7u) : 0u) +
-			    (emit_gsimg ? (9u + ((gs_mode & 1u) ? 11u : 0u)) : 0u) +
+			    (emit_gsimg ? (9u + ((gs_mode & 1u) ? 11u : 0u) +
+					   ((gs_ctx & 1u) ? 2u : 0u) +
+					   ((gs_ctx & 2u) ? 2u : 0u)) : 0u) +
 			    (emit_raster ? 17u : 0u)
 			    <= 96u) {
 				/*
@@ -30211,7 +30251,8 @@ gr_page_unmapped:
 				 * run here (nrun++ below).
 				 */
 				if (emit_gsimg) {
-					u64 gva = fam + 0x780u;
+					u64 gva = fam + (gs_addr ? 0xe80u :
+							 0x780u);
 
 					in[nfin++] = 0x20000000u | (1u << 16) |
 						     (0x2100u >> 2);
@@ -30246,6 +30287,18 @@ gr_page_unmapped:
 						in[nfin++] = 0u;
 						in[nfin++] = 0u;
 					}
+					if (gs_ctx & 1u) {
+						in[nfin++] = 0x20000000u |
+							     (1u << 16) |
+							     (0x114cu >> 2);
+						in[nfin++] = 0u;
+					}
+					if (gs_ctx & 2u) {
+						in[nfin++] = 0x20000000u |
+							     (1u << 16) |
+							     (0x0208u >> 2);
+						in[nfin++] = 0u;
+					}
 					in[nfin++] = 0x20000000u | (1u << 16) |
 						     (0x2100u >> 2);
 					in[nfin++] = (gs_mode & 2u) ? 0x41u :
diff --git a/kernel/gsp_phys_gpu.h b/kernel/gsp_phys_gpu.h
index d7177a6..d84cd41 100644
--- a/kernel/gsp_phys_gpu.h
+++ b/kernel/gsp_phys_gpu.h
@@ -375,6 +375,12 @@ u32 gsp_phys_rmcall_gr_sph_gs_mode_value(void);
 bool gsp_phys_rmcall_gr_raster_pre_value(void);
 
 /*
+ * SESSION 155 CONT 15: synthesized GS SPH placement/maps/context A/B.
+ */
+u32 gsp_phys_rmcall_gr_sph_gs_addr_value(void);
+u32 gsp_phys_rmcall_gr_sph_gs_ctx_value(void);
+
+/*
  * r1512: the viewport state the 3D replay injects for the current launch
  * (sema_gl_vp_prelude mode + its literal words). Read-only; used by the
  * sema service's draw-coverage receipt.
diff --git a/kernel/gsp_sema_service.c b/kernel/gsp_sema_service.c
index 4ede0c7..6972172 100644
--- a/kernel/gsp_sema_service.c
+++ b/kernel/gsp_sema_service.c
@@ -10657,7 +10657,9 @@ static void svc_gl_flush(struct nv_vgpu_device *vgpu,
 				};
 				u32 gs_maps =
 					gsp_phys_rmcall_gr_sph_gs_maps_value();
-				u64 dst = fam + 0x780u;
+				u64 dst = fam +
+					(gsp_phys_rmcall_gr_sph_gs_addr_value() ?
+					 0xe80u : 0x780u);
 				bool dup = false;
 				u32 pi4;
 
@@ -10678,6 +10680,19 @@ static void svc_gl_flush(struct nv_vgpu_device *vgpu,
 					gs_words[17] = 0u;
 				} else if (gs_maps == 3u) {
 					gs_words[17] = 0u;
+				} else if (gs_maps == 4u) {
+					/* IMAP = previous OMAP; OMAP =
+					 * PS IMAP exactly (POSITION.W). */
+					gs_words[5] = 0xf0000000u;
+					gs_words[10] = 0x0000000fu;
+					gs_words[13] = 0x00008000u;
+					gs_words[17] = 0u;
+				} else if (gs_maps == 5u) {
+					/* mesh-dummy IMAP, OMAP POSITION.W. */
+					gs_words[5] = 0u;
+					gs_words[10] = 0u;
+					gs_words[13] = 0x00008000u;
+					gs_words[17] = 0u;
 				}
 				/* SESSION 155 CONT 13: mesh-dummy FASTGS
 				 * form (bit24 ENABLE_FAST_G_S). */