vivante: fix composite using non-alpha pictures

The Vivante GPU hardware appears to be buggy with some formats and
mask combinations, and is particularly noticable with BGRA/RGBA
swizzles with non-alpha formats.  Fix this by switching to alpha
formats, and subsituting a maximal global alpha value instead.  This
allows rendercheck to pass for all formats with PE20 hardware.

Signed-off-by: Russell King <rmk@arm.linux.org.uk>

1 files changed, 108 insertions, 0 deletions

diff --git a/src/vivante_accel.c b/src/vivante_accel.c
index ae4fb3a..dc05b5f 100644
--- a/src/vivante_accel.c
+++ b/src/vivante_accel.c
@@ -1451,6 +1451,63 @@ int vivante_accel_Composite(CARD8 op, PicturePtr pSrc, PicturePtr pMask,
 
 	final_op = vivante_composite_op[op];
 
+	/*
+	 * There is a bug in the GPU hardware with destinations lacking
+	 * alpha and swizzles BGRA/RGBA.  Rather than the GPU treating
+	 * bits 7:0 as alpha, it continues to treat bits 31:24 as alpha.
+	 * This results in it replacing the B or R bits on input to the
+	 * blend operation with 1.0.  However, it continues to accept the
+	 * non-existent source alpha from bits 31:24.
+	 *
+	 * Work around this by switching to the equivalent alpha format,
+	 * and use global alpha to replace the alpha channel.
+	 */
+	if (!PICT_FORMAT_A(pDst->format)) {
+		final_op.dst_global_alpha = gcvSURF_GLOBAL_ALPHA_ON;
+		final_op.dst_alpha = 255;
+
+		switch (vDst->pict_format) {
+		case gcvSURF_X4R4G4B4:
+			vDst->pict_format = gcvSURF_A4R4G4B4;
+			break;
+		case gcvSURF_X4B4G4R4:
+			vDst->pict_format = gcvSURF_A4B4G4R4;
+			break;
+		case gcvSURF_R4G4B4X4:
+			vDst->pict_format = gcvSURF_R4G4B4A4;
+			break;
+		case gcvSURF_B4G4R4X4:
+			vDst->pict_format = gcvSURF_B4G4R4A4;
+			break;
+		case gcvSURF_X1R5G5B5:
+			vDst->pict_format = gcvSURF_A1R5G5B5;
+			break;
+		case gcvSURF_X1B5G5R5:
+			vDst->pict_format = gcvSURF_A1B5G5R5;
+			break;
+		case gcvSURF_R5G5B5X1:
+			vDst->pict_format = gcvSURF_R5G5B5A1;
+			break;
+		case gcvSURF_B5G5R5X1:
+			vDst->pict_format = gcvSURF_B5G5R5A1;
+			break;
+		case gcvSURF_X8R8G8B8:
+			vDst->pict_format = gcvSURF_A8R8G8B8;
+			break;
+		case gcvSURF_X8B8G8R8:
+			vDst->pict_format = gcvSURF_A8B8G8R8;
+			break;
+		case gcvSURF_R8G8B8X8:
+			vDst->pict_format = gcvSURF_R8G8B8A8;
+			break;
+		case gcvSURF_B8G8R8X8:
+			vDst->pict_format = gcvSURF_B8G8R8A8;
+			break;
+		default:
+			break;
+		}
+	}
+
 	if (pMask) {
 		uint32_t colour;
 
@@ -1613,6 +1670,57 @@ fprintf(stderr, "%s: i: op 0x%02x src=%p,%d,%d mask=%p,%d,%d dst=%p,%d,%d %ux%u\
 					   &xSrc, &ySrc);
 		if (!vSrc)
 			goto failed;
+
+		/*
+		 * Apply the same work-around for a non-alpha source as for
+		 * a non-alpha destination.
+		 */
+		if (!pMask && vSrc != vTemp && !PICT_FORMAT_A(pSrc->format) &&
+		    final_op.src_global_alpha == gcvSURF_GLOBAL_ALPHA_OFF) {
+			final_op.src_global_alpha = gcvSURF_GLOBAL_ALPHA_ON;
+			final_op.src_alpha = 255;
+
+			switch (vSrc->pict_format) {
+			case gcvSURF_X4R4G4B4:
+				vSrc->pict_format = gcvSURF_A4R4G4B4;
+				break;
+			case gcvSURF_X4B4G4R4:
+				vSrc->pict_format = gcvSURF_A4B4G4R4;
+				break;
+			case gcvSURF_R4G4B4X4:
+				vSrc->pict_format = gcvSURF_R4G4B4A4;
+				break;
+			case gcvSURF_B4G4R4X4:
+				vSrc->pict_format = gcvSURF_B4G4R4A4;
+				break;
+			case gcvSURF_X1R5G5B5:
+				vSrc->pict_format = gcvSURF_A1R5G5B5;
+				break;
+			case gcvSURF_X1B5G5R5:
+				vSrc->pict_format = gcvSURF_A1B5G5R5;
+				break;
+			case gcvSURF_R5G5B5X1:
+				vSrc->pict_format = gcvSURF_R5G5B5A1;
+				break;
+			case gcvSURF_B5G5R5X1:
+				vSrc->pict_format = gcvSURF_B5G5R5A1;
+				break;
+			case gcvSURF_X8R8G8B8:
+				vSrc->pict_format = gcvSURF_A8R8G8B8;
+				break;
+			case gcvSURF_X8B8G8R8:
+				vSrc->pict_format = gcvSURF_A8B8G8R8;
+				break;
+			case gcvSURF_R8G8B8X8:
+				vSrc->pict_format = gcvSURF_R8G8B8A8;
+				break;
+			case gcvSURF_B8G8R8X8:
+				vSrc->pict_format = gcvSURF_B8G8R8A8;
+				break;
+			default:
+				break;
+			}
+		}
 	}
 
 //vivante_batch_wait_commit(vivante, vSrc);