X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fpanfrost%2Fbifrost%2Fbifrost.h;h=12fb13a2c0c432be2ffe3f3f0d0fda0ee3646c90;hb=3e7fd2a0346ee364b1b0ccf58cc9214a602fcc35;hp=3e8e0229e9d0f43638c7c900aec5a47039a8eecd;hpb=07b31f3437ef60779f0fea83425521da3b7441f9;p=mesa.git

diff --git a/src/panfrost/bifrost/bifrost.h b/src/panfrost/bifrost/bifrost.h
index 3e8e0229e9d..12fb13a2c0c 100644
--- a/src/panfrost/bifrost/bifrost.h
+++ b/src/panfrost/bifrost/bifrost.h
@@ -42,6 +42,7 @@ enum bifrost_clause_type {
         BIFROST_CLAUSE_SSBO_LOAD  = 5,
         BIFROST_CLAUSE_SSBO_STORE = 6,
         BIFROST_CLAUSE_BLEND      = 9,
+        BIFROST_CLAUSE_FRAGZ      = 12,
         BIFROST_CLAUSE_ATEST      = 13,
         BIFROST_CLAUSE_64BIT      = 15
 };
@@ -114,6 +115,10 @@ struct bifrost_fma_inst {
         unsigned op   : 20;
 } __attribute__((packed));
 
+#define BIFROST_FMA_IADD_32 (0x4ff98 >> 3)
+#define BIFROST_FMA_ISUB_32 (0x4ffd8 >> 3)
+#define BIFROST_FMA_IMUL_32 ((BIFROST_FMA_EXT | 0x7818) >> 3)
+
 struct bifrost_fma_2src {
         unsigned src0 : 3;
         unsigned src1 : 3;
@@ -158,19 +163,47 @@ struct bifrost_fma_mscale {
 #define BIFROST_ADD_OP_FRSQ_FAST_F16_X (0x0ce50)
 #define BIFROST_ADD_OP_FRSQ_FAST_F16_Y (0x0ce70)
 #define BIFROST_ADD_OP_LOG2_HELP  (0x0cc68)
-#define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58)
+#define BIFROST_ADD_OP_IABS_32 (0x07bd4)
 
 struct bifrost_add_inst {
         unsigned src0 : 3;
         unsigned op   : 17;
 } __attribute__((packed));
 
+#define BIFROST_ADD_OP_DISCARD (0x19100 >> 8)
+
+enum bifrost_discard_cond {
+        BIFROST_DISCARD_FEQ = 0,
+        BIFROST_DISCARD_FNE = 1,
+        BIFROST_DISCARD_FLE = 2,
+        BIFROST_DISCARD_FLT = 3,
+};
+
+struct bifrost_add_discard {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        enum bifrost_discard_cond cond : 2;
+        /* Zero for fp32 */
+        unsigned src0_select : 1;
+        unsigned src1_select : 1;
+        unsigned fp32 : 1;
+        unsigned op   : 9;
+} __attribute__((packed));
+
 #define BIFROST_ADD_OP_LD_UBO_1 (0x0c1a0 >> 3)
 #define BIFROST_ADD_OP_LD_UBO_2 (0x0c1e0 >> 3)
 #define BIFROST_ADD_OP_LD_UBO_3 (0x0caa0 >> 3)
 #define BIFROST_ADD_OP_LD_UBO_4 (0x0c220 >> 3)
 #define BIFROST_ADD_SEL_16(swiz) ((0xea60 >> 3) | (swiz))
 
+#define BIFROST_ADD_IADD_8  (0x17880 >> 3)
+#define BIFROST_ADD_IADD_16 (0x17900 >> 3)
+#define BIFROST_ADD_IADD_32 (0x178c0 >> 3)
+#define BIFROST_ADD_ISUB_8  (0x17a80 >> 3)
+#define BIFROST_ADD_ISUB_16 (0x17b00 >> 3)
+#define BIFROST_ADD_ISUB_32 (0x17ac0 >> 3)
+#define BIFROST_ADD_OP_FEXP2_FAST (0x0cd58 >> 3)
+
 struct bifrost_add_2src {
         unsigned src0 : 3;
         unsigned src1 : 3;
@@ -416,6 +449,7 @@ enum bifrost_fcmp_cond {
         BIFROST_OLE = 5,
 };
 
+/* "gl" version produces 0/1. "d3d" version produces 0/~0 */
 #define BIFROST_FMA_OP_FCMP_GL (0x48000 >> 13)
 #define BIFROST_FMA_OP_FCMP_D3D (0x4c000 >> 13)
 
@@ -442,6 +476,7 @@ struct bifrost_add_fcmp {
         unsigned op   : 6;
 } __attribute__((packed));
 
+/* "gl" version produces 0/1. "d3d" version produces 0/~0 */
 #define BIFROST_FMA_OP_FCMP_GL_16 (0xc8000 >> 13)
 #define BIFROST_FMA_OP_FCMP_D3D_16 (0xcc000 >> 13)
 
@@ -476,11 +511,13 @@ struct bifrost_add_fcmp16 {
 
 enum bifrost_icmp_cond {
         BIFROST_ICMP_IGT = 0,
-        BIFROST_ICMP_IGE = 1,
-        BIFROST_ICMP_UGT = 2,
+        BIFROST_ICMP_IGE = 1, /* swapped for 16-bit */
+        BIFROST_ICMP_UGT = 2, /* swapped for 16-bit */
         BIFROST_ICMP_UGE = 3,
         BIFROST_ICMP_EQ  = 4,
         BIFROST_ICMP_NEQ  = 5,
+        BIFROST_ICMP_32_OR_8 = 6, /* nested */
+        BIFROST_ICMP_64 = 7, /* nested */
 };
 
 struct bifrost_fma_icmp32 {
@@ -488,7 +525,7 @@ struct bifrost_fma_icmp32 {
         unsigned src1 : 3;
         enum bifrost_icmp_cond cond : 3;
         unsigned unk1 : 1; /* set */
-        unsigned d3d : 1;
+        unsigned d3d : 1;  /* if set, true is ~0. otherwise, true is 1 */
         unsigned op : 12;
 } __attribute__((packed));
 
@@ -500,15 +537,28 @@ struct bifrost_fma_icmp16 {
         unsigned op : 9;
 } __attribute__((packed));
 
+#define BIFROST_ADD_OP_ICMP_32 (0x0f600 >> 8)
+#define BIFROST_ADD_OP_ICMP_16 (0x0f000 >> 11)
+
 struct bifrost_add_icmp {
         unsigned src0 : 3;
         unsigned src1 : 3;
         enum bifrost_icmp_cond cond : 3;
         unsigned sz : 1; /* 1 for 32, 0 for 8 */
-        unsigned d3d : 1;
+        unsigned d3d : 1; /* if set, true is ~0. otherwise, true is 1 */
         unsigned op : 9;
 } __attribute__((packed));
 
+struct bifrost_add_icmp16 {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        unsigned src0_swizzle : 2;
+        unsigned src1_swizzle : 2;
+        unsigned d3d : 1; /* if set, true is ~0. otherwise, true is 1 */
+        enum bifrost_icmp_cond cond : 3;
+        unsigned op : 6;
+} __attribute__((packed));
+ 
 /* Two sources for vectorization */
 #define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
 #define BIFROST_ADD_FLOAT32_TO_16 (0x0EC00 >> 3)
@@ -578,6 +628,10 @@ enum bifrost_interp_mode {
 #define BIFROST_ADD_OP_LD_VAR_16 (0x1a << 1)
 #define BIFROST_ADD_OP_LD_VAR_32 (0x0a << 1)
 
+/* Fixed location for gl_FragCoord.zw */
+#define BIFROST_FRAGZ (23)
+#define BIFROST_FRAGW (22)
+
 struct bifrost_ld_var {
         unsigned src0 : 3;
 
@@ -618,14 +672,14 @@ struct bifrost_dual_tex_ctrl {
         unsigned unk1 : 22;
 } __attribute__((packed));
 
-#define BIFROST_ADD_OP_TEX_COMPACT_F32 (0x0b000 >> 10)
-#define BIFROST_ADD_OP_TEX_COMPACT_F16 (0x1b000 >> 10)
+#define BIFROST_ADD_OP_TEX_COMPACT_F32(vtx) ((0x0b000 | ((vtx) ? (0x400) : (0))) >> 10)
+#define BIFROST_ADD_OP_TEX_COMPACT_F16(vtx) ((0x1b000 | ((vtx) ? (0x400) : (0))) >> 10)
 
 struct bifrost_tex_compact {
         unsigned src0 : 3;
         unsigned src1 : 3;
         unsigned tex_index : 3;
-        unsigned unknown : 1;
+        unsigned compute_lod : 1;
         unsigned sampler_index : 3;
         unsigned op   : 7;
 } __attribute__((packed));
@@ -684,6 +738,8 @@ enum bifrost_branch_code {
         BR_ALWAYS = 63,
 };
 
+#define BIFROST_ADD_OP_BRANCH (0x0d000 >> 12)
+
 struct bifrost_branch {
         unsigned src0 : 3;