drm/radeon: group r6xx/r7xx sequential blit state
group state that is emitted sequentially into fewer packets.
This saves a number of dwords.
Signed-off-by: Alex Deucher <alexdeucher@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
diff --git a/drivers/gpu/drm/radeon/r600_blit_shaders.c b/drivers/gpu/drm/radeon/r600_blit_shaders.c
index 3bcbb93..3a775c1 100644
--- a/drivers/gpu/drm/radeon/r600_blit_shaders.c
+++ b/drivers/gpu/drm/radeon/r600_blit_shaders.c
@@ -70,12 +70,9 @@
0x0000060e,
0x01020204, /* DB_WATERMARKS */
- 0xc0016f00,
+ 0xc0026f00,
0x00000000,
0x00000000, /* SQ_VTX_BASE_VTX_LOC */
-
- 0xc0016f00,
- 0x00000001,
0x00000000, /* SQ_VTX_START_INST_LOC */
0xc0096900,
@@ -94,32 +91,23 @@
0x00000004,
0x00000000, /* DB_DEPTH_INFO */
- 0xc0016900,
+ 0xc0026900,
0x0000000a,
0x00000000, /* DB_STENCIL_CLEAR */
-
- 0xc0016900,
- 0x0000000b,
0x00000000, /* DB_DEPTH_CLEAR */
- 0xc0016900,
+ 0xc0026900,
0x0000010c,
0x00000000, /* DB_STENCILREFMASK */
-
- 0xc0016900,
- 0x0000010d,
0x00000000, /* DB_STENCILREFMASK_BF */
0xc0016900,
0x00000200,
0x00000000, /* DB_DEPTH_CONTROL */
- 0xc0016900,
+ 0xc0026900,
0x00000343,
0x00000060, /* DB_RENDER_CONTROL */
-
- 0xc0016900,
- 0x00000344,
0x00000040, /* DB_RENDER_OVERRIDE */
0xc0016900,
@@ -134,15 +122,12 @@
0x0000010e,
0x00000000, /* SX_ALPHA_REF */
- 0xc0046900,
+ 0xc0076900,
0x00000105,
0x00000000, /* CB_BLEND_RED */
0x00000000,
0x00000000,
0x00000000,
-
- 0xc0036900,
- 0x00000109,
0x00000000, /* CB_FOG_RED */
0x00000000,
0x00000000,
@@ -169,52 +154,22 @@
0x00000080,
0x00000000, /* PA_SC_WINDOW_OFFSET */
- 0xc0016900,
+ 0xc00a6900,
0x00000083,
0x0000ffff, /* PA_SC_CLIP_RECT_RULE */
-
- 0xc0016900,
- 0x00000084,
- 0x00000000, /* PA_SC_WINDOW_SCISSOR_TL */
-
- 0xc0016900,
- 0x00000085,
+ 0x00000000, /* PA_SC_CLIPRECT_0_TL */
0x20002000,
-
- 0xc0016900,
- 0x00000086,
0x00000000,
-
- 0xc0016900,
- 0x00000087,
0x20002000,
-
- 0xc0016900,
- 0x00000088,
0x00000000,
-
- 0xc0016900,
- 0x00000089,
0x20002000,
-
- 0xc0016900,
- 0x0000008a,
0x00000000,
-
- 0xc0016900,
- 0x0000008b,
0x20002000,
-
- 0xc0016900,
- 0x0000008c,
0x00000000, /* PA_SC_EDGERULE */
- 0xc0016900,
+ 0xc0026900,
0x00000094,
0x80000000, /* PA_SC_VPORT_SCISSOR_0_TL */
-
- 0xc0016900,
- 0x00000095,
0x20002000, /* PA_SC_VPORT_SCISSOR_0_BR */
0xc0026900,
@@ -222,12 +177,9 @@
0x00000000, /* PA_SC_VPORT_ZMIN_0 */
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x00000096,
0x80000000, /* PA_SC_VPORT_SCISSOR_1_TL */
-
- 0xc0016900,
- 0x00000097,
0x20002000,
0xc0026900,
@@ -235,12 +187,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x00000098,
0x80000000,
-
- 0xc0016900,
- 0x00000099,
0x20002000,
0xc0026900,
@@ -248,12 +197,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x0000009a,
0x80000000,
-
- 0xc0016900,
- 0x0000009b,
0x20002000,
0xc0026900,
@@ -261,12 +207,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x0000009c,
0x80000000,
-
- 0xc0016900,
- 0x0000009d,
0x20002000,
0xc0026900,
@@ -274,12 +217,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x0000009e,
0x80000000,
-
- 0xc0016900,
- 0x0000009f,
0x20002000,
0xc0026900,
@@ -287,12 +227,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a0,
0x80000000,
-
- 0xc0016900,
- 0x000000a1,
0x20002000,
0xc0026900,
@@ -300,12 +237,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a2,
0x80000000,
-
- 0xc0016900,
- 0x000000a3,
0x20002000,
0xc0026900,
@@ -313,12 +247,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a4,
0x80000000,
-
- 0xc0016900,
- 0x000000a5,
0x20002000,
0xc0026900,
@@ -326,12 +257,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a6,
0x80000000,
-
- 0xc0016900,
- 0x000000a7,
0x20002000,
0xc0026900,
@@ -339,12 +267,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a8,
0x80000000,
-
- 0xc0016900,
- 0x000000a9,
0x20002000,
0xc0026900,
@@ -352,12 +277,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000aa,
0x80000000,
-
- 0xc0016900,
- 0x000000ab,
0x20002000,
0xc0026900,
@@ -365,12 +287,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000ac,
0x80000000,
-
- 0xc0016900,
- 0x000000ad,
0x20002000,
0xc0026900,
@@ -378,12 +297,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000ae,
0x80000000,
-
- 0xc0016900,
- 0x000000af,
0x20002000,
0xc0026900,
@@ -391,12 +307,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000b0,
0x80000000,
-
- 0xc0016900,
- 0x000000b1,
0x20002000,
0xc0026900,
@@ -404,12 +317,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000b2,
0x80000000,
-
- 0xc0016900,
- 0x000000b3,
0x20002000,
0xc0026900,
@@ -421,24 +331,18 @@
0x00000293,
0x00004010, /* PA_SC_MODE_CNTL */
- 0xc0016900,
+ 0xc0026900,
0x00000300,
0x00000000, /* PA_SC_LINE_CNTL */
-
- 0xc0016900,
- 0x00000301,
0x00000000, /* PA_SC_AA_CONFIG */
0xc0016900,
0x00000312,
0xffffffff, /* PA_SC_AA_MASK */
- 0xc0016900,
+ 0xc0026900,
0x00000307,
0x00000000, /* PA_SC_SAMPLE_LOCS_MCTX */
-
- 0xc0016900,
- 0x00000308,
0x00000000,
0xc0016900,
@@ -458,12 +362,9 @@
0x00000000,
0x00000000,
- 0xc0016900,
+ 0xc0026900,
0x00000207,
0x00000000, /* PA_CL_VS_OUT_CNTL */
-
- 0xc0016900,
- 0x00000208,
0x00000000, /* PA_CL_NANINF_CNTL */
0xc0046900,
@@ -473,12 +374,9 @@
0x3f800000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x00000280,
0x00000000, /* PA_SU_POINT_SIZE */
-
- 0xc0016900,
- 0x00000281,
0x00000000, /* PA_SU_POINT_MINMAX */
0xc0016900,
@@ -517,20 +415,11 @@
0x000001b2,
0x00000000, /* SPI_THREAD_GROUPING */
- 0xc0016900,
+ 0xc0046900,
0x000001b6,
0x00000000, /* SPI_INPUT_Z */
-
- 0xc0016900,
- 0x000001b7,
0x00000000, /* SPI_FOG_CNTL */
-
- 0xc0016900,
- 0x000001b8,
0x00000000, /* SPI_FOG_FUNC_SCALE */
-
- 0xc0016900,
- 0x000001b9,
0x00000000, /* SPI_FOG_FUNC_BIAS */
0xc0016900,
@@ -545,24 +434,15 @@
0x00000237,
0x00000000, /* SQ_PGM_CF_OFFSET_FS */
- 0xc0016900,
+ 0xc0036900,
0x00000100,
0x00000800, /* VGT_MAX_VTX_INDX */
-
- 0xc0016900,
- 0x00000101,
0x00000000, /* VGT_MIN_VTX_INDX */
-
- 0xc0016900,
- 0x00000102,
0x00000000, /* VGT_INDX_OFFSET */
- 0xc0016900,
+ 0xc0026900,
0x000002a8,
0x00000000, /* VGT_INSTANCE_STEP_RATE_0 */
-
- 0xc0016900,
- 0x000002a9,
0x00000000, /* VGT_INSTANCE_STEP_RATE_1 */
0xc0016900,
@@ -581,44 +461,17 @@
0x00000285,
0x00000000, /* VGT_HOS_CNTL */
- 0xc0016900,
+ 0xc00a6900,
0x00000286,
0x00000000, /* VGT_HOS_MAX_TESS_LEVEL */
-
- 0xc0016900,
- 0x00000287,
0x00000000, /* VGT_HOS_MIN_TESS_LEVEL */
-
- 0xc0016900,
- 0x00000288,
0x00000000, /* VGT_HOS_REUSE_DEPTH */
-
- 0xc0016900,
- 0x00000289,
0x00000000, /* VGT_GROUP_PRIM_TYPE */
-
- 0xc0016900,
- 0x0000028a,
0x00000000, /* VGT_GROUP_FIRST_DECR */
-
- 0xc0016900,
- 0x0000028b,
0x00000000, /* VGT_GROUP_DECR */
-
- 0xc0016900,
- 0x0000028c,
0x00000000, /* VGT_GROUP_VECT_0_CNTL */
-
- 0xc0016900,
- 0x0000028d,
0x00000000, /* VGT_GROUP_VECT_1_CNTL */
-
- 0xc0016900,
- 0x0000028e,
0x00000000, /* VGT_GROUP_VECT_0_FMT_CNTL */
-
- 0xc0016900,
- 0x0000028f,
0x00000000, /* VGT_GROUP_VECT_1_FMT_CNTL */
0xc0016900,
@@ -629,16 +482,10 @@
0x000002a5,
0x00000000, /* VGT_MULTI_PRIM_ID_RESET_EN */
- 0xc0016900,
+ 0xc0036900,
0x000002ac,
0x00000000, /* VGT_STRMOUT_EN */
-
- 0xc0016900,
- 0x000002ad,
0x00000000, /* VGT_REUSE_OFF */
-
- 0xc0016900,
- 0x000002ae,
0x00000000, /* VGT_VTX_CNT_EN */
0xc0016900,
@@ -687,12 +534,9 @@
0x00000185,
0x00000000, /* SPI_VS_OUT_ID_0 */
- 0xc0016900,
+ 0xc0026900,
0x000001b3,
0x00000001, /* SPI_PS_IN_CONTROL_0 */
-
- 0xc0016900,
- 0x000001b4,
0x00000000, /* SPI_PS_IN_CONTROL_1 */
0xc0016900,
@@ -734,12 +578,9 @@
0x0000060e,
0x00420204, /* DB_WATERMARKS */
- 0xc0016f00,
+ 0xc0026f00,
0x00000000,
0x00000000, /* SQ_VTX_BASE_VTX_LOC */
-
- 0xc0016f00,
- 0x00000001,
0x00000000, /* SQ_VTX_START_INST_LOC */
0xc0096900,
@@ -758,32 +599,23 @@
0x00000004,
0x00000000, /* DB_DEPTH_INFO */
- 0xc0016900,
+ 0xc0026900,
0x0000000a,
0x00000000, /* DB_STENCIL_CLEAR */
-
- 0xc0016900,
- 0x0000000b,
0x00000000, /* DB_DEPTH_CLEAR */
- 0xc0016900,
+ 0xc0026900,
0x0000010c,
0x00000000, /* DB_STENCILREFMASK */
-
- 0xc0016900,
- 0x0000010d,
0x00000000, /* DB_STENCILREFMASK_BF */
0xc0016900,
0x00000200,
0x00000000, /* DB_DEPTH_CONTROL */
- 0xc0016900,
+ 0xc0026900,
0x00000343,
0x00000060, /* DB_RENDER_CONTROL */
-
- 0xc0016900,
- 0x00000344,
0x00000000, /* DB_RENDER_OVERRIDE */
0xc0016900,
@@ -820,52 +652,22 @@
0x00000080,
0x00000000, /* PA_SC_WINDOW_OFFSET */
- 0xc0016900,
+ 0xc00a6900,
0x00000083,
0x0000ffff, /* PA_SC_CLIP_RECT_RULE */
-
- 0xc0016900,
- 0x00000084,
- 0x00000000, /* PA_SC_WINDOW_SCISSOR_TL */
-
- 0xc0016900,
- 0x00000085,
+ 0x00000000, /* PA_SC_CLIPRECT_0_TL */
0x20002000,
-
- 0xc0016900,
- 0x00000086,
0x00000000,
-
- 0xc0016900,
- 0x00000087,
0x20002000,
-
- 0xc0016900,
- 0x00000088,
0x00000000,
-
- 0xc0016900,
- 0x00000089,
0x20002000,
-
- 0xc0016900,
- 0x0000008a,
0x00000000,
-
- 0xc0016900,
- 0x0000008b,
0x20002000,
-
- 0xc0016900,
- 0x0000008c,
0xaaaaaaaa, /* PA_SC_EDGERULE */
- 0xc0016900,
+ 0xc0026900,
0x00000094,
0x80000000, /* PA_SC_VPORT_SCISSOR_0_TL */
-
- 0xc0016900,
- 0x00000095,
0x20002000, /* PA_SC_VPORT_SCISSOR_0_BR */
0xc0026900,
@@ -873,12 +675,9 @@
0x00000000, /* PA_SC_VPORT_ZMIN_0 */
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x00000096,
0x80000000,
-
- 0xc0016900,
- 0x00000097,
0x20002000,
0xc0026900,
@@ -886,12 +685,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x00000098,
0x80000000,
-
- 0xc0016900,
- 0x00000099,
0x20002000,
0xc0026900,
@@ -902,9 +698,6 @@
0xc0016900,
0x0000009a,
0x80000000,
-
- 0xc0016900,
- 0x0000009b,
0x20002000,
0xc0026900,
@@ -912,12 +705,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x0000009c,
0x80000000,
-
- 0xc0016900,
- 0x0000009d,
0x20002000,
0xc0026900,
@@ -925,12 +715,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x0000009e,
0x80000000,
-
- 0xc0016900,
- 0x0000009f,
0x20002000,
0xc0026900,
@@ -938,12 +725,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a0,
0x80000000,
-
- 0xc0016900,
- 0x000000a1,
0x20002000,
0xc0026900,
@@ -951,12 +735,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a2,
0x80000000,
-
- 0xc0016900,
- 0x000000a3,
0x20002000,
0xc0026900,
@@ -964,12 +745,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a4,
0x80000000,
-
- 0xc0016900,
- 0x000000a5,
0x20002000,
0xc0026900,
@@ -977,12 +755,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a6,
0x80000000,
-
- 0xc0016900,
- 0x000000a7,
0x20002000,
0xc0026900,
@@ -990,12 +765,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000a8,
0x80000000,
-
- 0xc0016900,
- 0x000000a9,
0x20002000,
0xc0026900,
@@ -1003,12 +775,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000aa,
0x80000000,
-
- 0xc0016900,
- 0x000000ab,
0x20002000,
0xc0026900,
@@ -1016,12 +785,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000ac,
0x80000000,
-
- 0xc0016900,
- 0x000000ad,
0x20002000,
0xc0026900,
@@ -1029,12 +795,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000ae,
0x80000000,
-
- 0xc0016900,
- 0x000000af,
0x20002000,
0xc0026900,
@@ -1042,12 +805,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000b0,
0x80000000,
-
- 0xc0016900,
- 0x000000b1,
0x20002000,
0xc0026900,
@@ -1055,12 +815,9 @@
0x00000000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x000000b2,
0x80000000,
-
- 0xc0016900,
- 0x000000b3,
0x20002000,
0xc0026900,
@@ -1072,24 +829,18 @@
0x00000293,
0x00514000, /* PA_SC_MODE_CNTL */
- 0xc0016900,
+ 0xc0026900,
0x00000300,
0x00000000, /* PA_SC_LINE_CNTL */
-
- 0xc0016900,
- 0x00000301,
0x00000000, /* PA_SC_AA_CONFIG */
0xc0016900,
0x00000312,
0xffffffff, /* PA_SC_AA_MASK */
- 0xc0016900,
+ 0xc0026900,
0x00000307,
0x00000000, /* PA_SC_SAMPLE_LOCS_MCTX */
-
- 0xc0016900,
- 0x00000308,
0x00000000,
0xc0016900,
@@ -1109,12 +860,9 @@
0x00000000,
0x00000000,
- 0xc0016900,
+ 0xc0026900,
0x00000207,
0x00000000, /* PA_CL_VS_OUT_CNTL */
-
- 0xc0016900,
- 0x00000208,
0x00000000, /* PA_CL_NANINF_CNTL */
0xc0046900,
@@ -1124,12 +872,9 @@
0x3f800000,
0x3f800000,
- 0xc0016900,
+ 0xc0026900,
0x00000280,
0x00000000, /* PA_SU_POINT_SIZE */
-
- 0xc0016900,
- 0x00000281,
0x00000000, /* PA_SU_POINT_MINMAX */
0xc0016900,
@@ -1168,20 +913,11 @@
0x000001b2,
0x00000001, /* SPI_THREAD_GROUPING */
- 0xc0016900,
+ 0xc0046900,
0x000001b6,
0x00000000, /* SPI_INPUT_Z */
-
- 0xc0016900,
- 0x000001b7,
0x00000000, /* SPI_FOG_CNTL */
-
- 0xc0016900,
- 0x000001b8,
0x00000000, /* SPI_FOG_FUNC_SCALE */
-
- 0xc0016900,
- 0x000001b9,
0x00000000, /* SPI_FOG_FUNC_BIAS */
0xc0016900,
@@ -1196,24 +932,15 @@
0x00000237,
0x00000000, /* SQ_PGM_CF_OFFSET_FS */
- 0xc0016900,
+ 0xc0036900,
0x00000100,
0x00000800, /* VGT_MAX_VTX_INDX */
-
- 0xc0016900,
- 0x00000101,
0x00000000, /* VGT_MIN_VTX_INDX */
-
- 0xc0016900,
- 0x00000102,
0x00000000, /* VGT_INDX_OFFSET */
- 0xc0016900,
+ 0xc0026900,
0x000002a8,
0x00000000, /* VGT_INSTANCE_STEP_RATE_0 */
-
- 0xc0016900,
- 0x000002a9,
0x00000000, /* VGT_INSTANCE_STEP_RATE_1 */
0xc0016900,
@@ -1228,48 +955,18 @@
0x00000290,
0x00000000, /* VGT_GS_MODE */
- 0xc0016900,
+ 0xc00b6900,
0x00000285,
0x00000000, /* VGT_HOS_CNTL */
-
- 0xc0016900,
- 0x00000286,
0x00000000, /* VGT_HOS_MAX_TESS_LEVEL */
-
- 0xc0016900,
- 0x00000287,
0x00000000, /* VGT_HOS_MIN_TESS_LEVEL */
-
- 0xc0016900,
- 0x00000288,
0x00000000, /* VGT_HOS_REUSE_DEPTH */
-
- 0xc0016900,
- 0x00000289,
0x00000000, /* VGT_GROUP_PRIM_TYPE */
-
- 0xc0016900,
- 0x0000028a,
0x00000000, /* VGT_GROUP_FIRST_DECR */
-
- 0xc0016900,
- 0x0000028b,
0x00000000, /* VGT_GROUP_DECR */
-
- 0xc0016900,
- 0x0000028c,
0x00000000, /* VGT_GROUP_VECT_0_CNTL */
-
- 0xc0016900,
- 0x0000028d,
0x00000000, /* VGT_GROUP_VECT_1_CNTL */
-
- 0xc0016900,
- 0x0000028e,
0x00000000, /* VGT_GROUP_VECT_0_FMT_CNTL */
-
- 0xc0016900,
- 0x0000028f,
0x00000000, /* VGT_GROUP_VECT_1_FMT_CNTL */
0xc0016900,
@@ -1280,16 +977,10 @@
0x000002a5,
0x00000000, /* VGT_MULTI_PRIM_ID_RESET_EN */
- 0xc0016900,
+ 0xc0036900,
0x000002ac,
0x00000000, /* VGT_STRMOUT_EN */
-
- 0xc0016900,
- 0x000002ad,
0x00000000, /* VGT_REUSE_OFF */
-
- 0xc0016900,
- 0x000002ae,
0x00000000, /* VGT_VTX_CNT_EN */
0xc0016900,
@@ -1338,12 +1029,9 @@
0x00000185,
0x00000000, /* SPI_VS_OUT_ID_0 */
- 0xc0016900,
+ 0xc0026900,
0x000001b3,
0x00000001, /* SPI_PS_IN_CONTROL_0 */
-
- 0xc0016900,
- 0x000001b4,
0x00000000, /* SPI_PS_IN_CONTROL_1 */
0xc0016900,