X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fpanfrost%2Fbifrost%2Fbifrost.h;h=12fb13a2c0c432be2ffe3f3f0d0fda0ee3646c90;hb=HEAD;hp=31ea14db63c42d66c1ecfa18575083a078b980ac;hpb=d7bb7b79a8b0a86ee3fbd21fffa8e3d0c8dd03a2;p=mesa.git

diff --git a/src/panfrost/bifrost/bifrost.h b/src/panfrost/bifrost/bifrost.h
index 31ea14db63c..12fb13a2c0c 100644
--- a/src/panfrost/bifrost/bifrost.h
+++ b/src/panfrost/bifrost/bifrost.h
@@ -29,6 +29,11 @@
 #include <stdint.h>
 #include <stdbool.h>
 
+#define BIFROST_DBG_MSGS        0x0001
+#define BIFROST_DBG_SHADERS     0x0002
+
+extern int bifrost_debug;
+
 enum bifrost_clause_type {
         BIFROST_CLAUSE_NONE       = 0,
         BIFROST_CLAUSE_LOAD_VARY  = 1,
@@ -37,6 +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
 };
@@ -97,31 +103,106 @@ enum bifrost_packed_src {
 
 #define BIFROST_FMA_EXT (0xe0000)
 #define BIFROST_FMA_OP_MOV BIFROST_FMA_EXT | (0x32d)
+#define BIFROST_FMA_OP_FREXPE_LOG BIFROST_FMA_EXT | 0x3c5
+#define BIFROST_FMA_OP_ADD_FREXPM ((BIFROST_FMA_EXT | 0x1e80) >> 3)
+#define BIFROST_FMA_SEL_16(swiz) (((BIFROST_FMA_EXT | 0x1e00) >> 3) | (swiz))
+
+#define BIFROST_FMA_ROUND_16(mode, swiz) (BIFROST_FMA_EXT | 0x1800 | (swiz) | ((mode) << 6))
+#define BIFROST_FMA_ROUND_32(mode) (BIFROST_FMA_EXT | 0x1805 | ((mode) << 6))
 
 struct bifrost_fma_inst {
         unsigned src0 : 3;
         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;
         unsigned op   : 17;
 } __attribute__((packed));
 
+#define BIFROST_FMA_OP_SEL8 (0x71)
+
+struct bifrost_fma_sel8 {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        unsigned src2 : 3;
+        unsigned src3 : 3;
+        unsigned swizzle : 4;
+        unsigned op   : 7;
+} __attribute__((packed));
+
+#define BIFROST_FMA_OP_MSCALE (0x50 >> 3)
+
+struct bifrost_fma_mscale {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        unsigned src2 : 3;
+        unsigned src3 : 3;
+
+        /* If mscale_mode is set - an MSCALE specific mode. If it is not set, a
+         * regular outmod */
+        unsigned mode : 2;
+        unsigned mscale_mode : 1;
+
+        unsigned src0_abs : 1;
+        unsigned src1_neg : 1;
+        unsigned src2_neg : 1;
+        unsigned op   : 5;
+} __attribute__((packed));
+
 #define BIFROST_ADD_OP_BLEND (0x1952c)
 #define BIFROST_ADD_OP_FRCP_FAST_F32 (0x0cc00)
+#define BIFROST_ADD_OP_FRCP_FAST_F16_X (0x0ce10)
+#define BIFROST_ADD_OP_FRCP_FAST_F16_Y (0x0ce30)
 #define BIFROST_ADD_OP_FRSQ_FAST_F32 (0x0cc20)
+#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_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;
@@ -129,6 +210,41 @@ struct bifrost_add_2src {
         unsigned op   : 14;
 } __attribute__((packed));
 
+#define BIFROST_ADD_OP_FMAX32 (0x00)
+#define BIFROST_ADD_OP_FMIN32 (0x01)
+#define BIFROST_ADD_OP_FADD32 (0x02)
+
+#define BIFROST_ADD_OP_FADD16 (0x0A)
+
+struct bifrost_add_faddmin {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        unsigned src1_abs : 1;
+        unsigned src0_neg : 1;
+        unsigned src1_neg : 1;
+        unsigned select : 2; /* swizzle_0 for fp16 */
+        unsigned outmod : 2; /* swizzle_1 for fp16 */
+        unsigned mode : 2;
+        unsigned src0_abs : 1;
+        unsigned op   : 4;
+} __attribute__((packed));
+
+#define BIFROST_ADD_OP_FMAX16 (0x10)
+#define BIFROST_ADD_OP_FMIN16 (0x12)
+
+struct bifrost_add_fmin16 {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        /* abs2 inferred as with FMA */
+        unsigned abs1 : 1;
+        unsigned src0_neg : 1;
+        unsigned src1_neg : 1;
+        unsigned src0_swizzle : 2;
+        unsigned src1_swizzle : 2;
+        unsigned mode : 2;
+        unsigned op : 5;
+} __attribute__((packed));
+
 #define BIFROST_ADD_OP_ST_VAR (0x19300 >> 8)
 
 struct bifrost_st_vary {
@@ -197,6 +313,8 @@ enum bifrost_minmax_mode {
 };
 
 #define BIFROST_FMA_OP_FADD32 (0x58 >> 2)
+#define BIFROST_FMA_OP_FMAX32 (0x40 >> 2)
+#define BIFROST_FMA_OP_FMIN32 (0x44 >> 2)
 
 struct bifrost_fma_add {
         unsigned src0 : 3;
@@ -286,11 +404,23 @@ struct bifrost_csel4 {
         unsigned op   : 8;
 } __attribute__((packed));
 
+#define BIFROST_FMA_OP_RSHIFT_NAND     (0x60000 >> 12)
+#define BIFROST_FMA_OP_RSHIFT_AND      (0x61000 >> 12)
+#define BIFROST_FMA_OP_LSHIFT_NAND     (0x62000 >> 12)
+#define BIFROST_FMA_OP_LSHIFT_AND      (0x63000 >> 12)
+#define BIFROST_FMA_OP_RSHIFT_XOR      (0x64000 >> 12)
+#define BIFROST_FMA_OP_LSHIFT_ADD_32   (0x65200 >> 6)
+#define BIFROST_FMA_OP_LSHIFT_SUB_32   (0x65600 >> 6)
+#define BIFROST_FMA_OP_LSHIFT_RSUB_32  (0x65a00 >> 6)
+#define BIFROST_FMA_OP_RSHIFT_ADD_32   (0x65e00 >> 6)
+#define BIFROST_FMA_OP_RSHIFT_SUB_32   (0x66200 >> 6)
+#define BIFROST_FMA_OP_RSHIFT_RSUB_32  (0x66600 >> 6)
+
 struct bifrost_shift_fma {
         unsigned src0 : 3;
         unsigned src1 : 3;
         unsigned src2 : 3;
-        unsigned half : 3; /* 000 for i32, 100 for i8, 111 for v2i16 */
+        unsigned half : 3;
         unsigned unk  : 1; /* always set? */
         unsigned invert_1 : 1; /* Inverts sources to combining op */
         /* For XOR, switches RSHIFT to LSHIFT since only one invert needed */
@@ -310,40 +440,156 @@ struct bifrost_shift_add {
         unsigned op : 7;
 } __attribute__((packed));
 
-#define BIFROST_FMA_INT16_TO_32 (0xe0198 >> 3)
+enum bifrost_fcmp_cond {
+        BIFROST_OEQ = 0,
+        BIFROST_OGT = 1,
+        BIFROST_OGE = 2,
+        BIFROST_UNE = 3,
+        BIFROST_OLT = 4,
+        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)
+
+struct bifrost_fma_fcmp {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        unsigned src1_abs : 1;
+        unsigned unk1 : 1;
+        unsigned src1_neg : 1;
+        unsigned src_expand : 3;
+        unsigned src0_abs : 1;
+        enum bifrost_fcmp_cond cond : 3;
+        unsigned op   : 7;
+} __attribute__((packed));
+
+struct bifrost_add_fcmp {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        enum bifrost_fcmp_cond cond : 3;
+        unsigned src_expand : 2;
+        unsigned src0_abs : 1;
+        unsigned src1_abs : 1;
+        unsigned src1_neg : 1;
+        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)
 
-struct bifrost_fma_int16_to_32 {
+struct bifrost_fma_fcmp16 {
         unsigned src0 : 3;
-        unsigned is_unsigned : 1;
-        unsigned swizzle : 1;
-        unsigned to_float : 1;
-        unsigned op : 17;
+        unsigned src1 : 3;
+
+        /* abs2 inferred */
+        unsigned abs1 : 1;
+        unsigned unk : 2;
+
+        unsigned src0_swizzle : 2;
+        unsigned src1_swizzle : 2;
+
+        enum bifrost_fcmp_cond cond : 3;
+        unsigned op   : 7;
 } __attribute__((packed));
 
-/* We could combine but it's easier to just use FMA_ONE_SRC */
-#define BIFROST_FMA_FLOAT16_TO_32(y) (0xe01a2 | ((y) ? 1 : 0))
+struct bifrost_add_fcmp16 {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        enum bifrost_fcmp_cond cond : 3;
 
-/* Two sources for vectorization */
-#define BIFROST_FMA_FLOAT32_TO_16 (0xdd000 >> 3)
+        unsigned src0_swizzle : 2;
+        unsigned src1_swizzle : 2;
 
-/* Again we could combine but easier to just ONE_SRC with an
- * argumnt for unsignedness */
-#define BIFROST_FMA_FLOAT32_TO_INT(u) (0xe0136 | ((u) ? 1 : 0))
-#define BIFROST_FMA_INT_TO_FLOAT32(u) (0xe0178 | ((u) ? 1 : 0))
+        /* No abs mods */
+        unsigned src0_neg : 1;
 
-/* Used for f2i16 and i2f16 */
-#define BIFROST_FMA_F2I16 (0xe00)
+        unsigned op   : 6;
+} __attribute__((packed));
+
+enum bifrost_icmp_cond {
+        BIFROST_ICMP_IGT = 0,
+        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_f2i_i2f16 {
+struct bifrost_fma_icmp32 {
         unsigned src0 : 3;
-        unsigned is_unsigned : 1;
-        unsigned direction : 2; /* 00 for i2f, 11 for f2i */
-        unsigned swizzle : 2;
-        unsigned unk : 2; /* always 10 */
-        unsigned direction_2 : 1; /* 0 for f2i, 1 for i2f */
+        unsigned src1 : 3;
+        enum bifrost_icmp_cond cond : 3;
+        unsigned unk1 : 1; /* set */
+        unsigned d3d : 1;  /* if set, true is ~0. otherwise, true is 1 */
         unsigned op : 12;
 } __attribute__((packed));
 
+struct bifrost_fma_icmp16 {
+        unsigned src0 : 3;
+        unsigned src1 : 3;
+        unsigned unk : 5; /* 11010 */
+        enum bifrost_icmp_cond cond : 3;
+        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; /* 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)
+
+enum bifrost_convert_mode {
+        BIFROST_CONV_UNK0 = 0,
+        BIFROST_CONV_F32_TO_I32 = 1,
+        BIFROST_CONV_F16_TO_I16 = 2,
+        BIFROST_CONV_I32_TO_F32 = 3,
+        BIFROST_CONV_I16_TO_X32 = 4,
+        BIFROST_CONV_F16_TO_F32 = 5,
+        BIFROST_CONV_I16_TO_F16 = 6,
+        BIFROST_CONV_UNK7 = 7
+};
+
+/* i16 to x32 */
+#define BIFROST_CONVERT_4(is_unsigned, component, to_float) \
+        ((is_unsigned & 1) | ((component & 1) << 1) | ((to_float & 1) << 2) | \
+         ((0x3) << 3) | ((4) << 5) | 0x100)
+
+/* f16 to f32 */
+#define BIFROST_CONVERT_5(component) \
+        ((component & 1) | ((1) << 1) | ((5) << 5) | 0x100)
+
+/* Other conversions */
+#define BIFROST_CONVERT(is_unsigned, roundmode, swizzle, mode) \
+        ((is_unsigned & 1) | ((roundmode & 3) << 1) | ((swizzle & 3) << 3) | ((mode & 7) << 5))
+
+#define BIFROST_FMA_CONVERT (0xe0000)
+#define BIFROST_ADD_CONVERT (0x07800)
+
 enum bifrost_ldst_type {
         BIFROST_LDST_F16 = 0,
         BIFROST_LDST_F32 = 1,
@@ -382,6 +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;
 
@@ -422,6 +672,18 @@ struct bifrost_dual_tex_ctrl {
         unsigned unk1 : 22;
 } __attribute__((packed));
 
+#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 compute_lod : 1;
+        unsigned sampler_index : 3;
+        unsigned op   : 7;
+} __attribute__((packed));
+
 enum branch_bit_size {
         BR_SIZE_32 = 0,
         BR_SIZE_16XX = 1,
@@ -476,6 +738,8 @@ enum bifrost_branch_code {
         BR_ALWAYS = 63,
 };
 
+#define BIFROST_ADD_OP_BRANCH (0x0d000 >> 12)
+
 struct bifrost_branch {
         unsigned src0 : 3;