#include <stdint.h>
#include <stdbool.h>
-#include <panfrost-misc.h>
-
-enum mali_job_type {
- JOB_NOT_STARTED = 0,
- JOB_TYPE_NULL = 1,
- JOB_TYPE_WRITE_VALUE = 2,
- JOB_TYPE_CACHE_FLUSH = 3,
- JOB_TYPE_COMPUTE = 4,
- JOB_TYPE_VERTEX = 5,
- JOB_TYPE_GEOMETRY = 6,
- JOB_TYPE_TILER = 7,
- JOB_TYPE_FUSED = 8,
- JOB_TYPE_FRAGMENT = 9,
-};
+#include <inttypes.h>
-enum mali_draw_mode {
- MALI_DRAW_NONE = 0x0,
- MALI_POINTS = 0x1,
- MALI_LINES = 0x2,
- MALI_LINE_STRIP = 0x4,
- MALI_LINE_LOOP = 0x6,
- MALI_TRIANGLES = 0x8,
- MALI_TRIANGLE_STRIP = 0xA,
- MALI_TRIANGLE_FAN = 0xC,
- MALI_POLYGON = 0xD,
- MALI_QUADS = 0xE,
- MALI_QUAD_STRIP = 0xF,
-
- /* All other modes invalid */
-};
+typedef uint8_t u8;
+typedef uint16_t u16;
+typedef uint32_t u32;
+typedef uint64_t u64;
+typedef uint64_t mali_ptr;
/* Applies to tiler_gl_enables */
#define MALI_CULL_FACE_FRONT (1 << 6)
#define MALI_CULL_FACE_BACK (1 << 7)
-/* Used in stencil and depth tests */
-
-enum mali_func {
- MALI_FUNC_NEVER = 0,
- MALI_FUNC_LESS = 1,
- MALI_FUNC_EQUAL = 2,
- MALI_FUNC_LEQUAL = 3,
- MALI_FUNC_GREATER = 4,
- MALI_FUNC_NOTEQUAL = 5,
- MALI_FUNC_GEQUAL = 6,
- MALI_FUNC_ALWAYS = 7
-};
-
/* Flags apply to unknown2_3? */
#define MALI_HAS_MSAA (1 << 0)
+
+/* Execute fragment shader per-sample if set (e.g. to implement gl_SampleID
+ * reads) */
+#define MALI_PER_SAMPLE (1 << 2)
#define MALI_CAN_DISCARD (1 << 5)
/* Applies on SFBD systems, specifying that programmable blending is in use */
#define MALI_DEPTH_WRITEMASK (1 << 11)
+#define MALI_DEPTH_CLIP_NEAR (1 << 12)
+#define MALI_DEPTH_CLIP_FAR (1 << 13)
+
/* Next flags to unknown2_4 */
#define MALI_STENCIL_TEST (1 << 0)
-/* What?! */
-#define MALI_SAMPLE_ALPHA_TO_COVERAGE_NO_BLEND_SHADER (1 << 1)
+#define MALI_ALPHA_TO_COVERAGE (1 << 1)
#define MALI_NO_DITHER (1 << 9)
#define MALI_DEPTH_RANGE_A (1 << 12)
#define MALI_DEPTH_RANGE_B (1 << 13)
#define MALI_NO_MSAA (1 << 14)
-/* Stencil test state is all encoded in a single u32, just with a lot of
- * enums... */
-
-enum mali_stencil_op {
- MALI_STENCIL_KEEP = 0,
- MALI_STENCIL_REPLACE = 1,
- MALI_STENCIL_ZERO = 2,
- MALI_STENCIL_INVERT = 3,
- MALI_STENCIL_INCR_WRAP = 4,
- MALI_STENCIL_DECR_WRAP = 5,
- MALI_STENCIL_INCR = 6,
- MALI_STENCIL_DECR = 7
-};
-
-struct mali_stencil_test {
- unsigned ref : 8;
- unsigned mask : 8;
- enum mali_func func : 3;
- enum mali_stencil_op sfail : 3;
- enum mali_stencil_op dpfail : 3;
- enum mali_stencil_op dppass : 3;
- unsigned zero : 4;
-} __attribute__((packed));
-
#define MALI_MASK_R (1 << 0)
#define MALI_MASK_G (1 << 1)
#define MALI_MASK_B (1 << 2)
MALI_CHANNEL_RESERVED_1 = 7,
};
-struct mali_channel_swizzle {
- enum mali_channel r : 3;
- enum mali_channel g : 3;
- enum mali_channel b : 3;
- enum mali_channel a : 3;
-} __attribute__((packed));
-
/* Compressed per-pixel formats. Each of these formats expands to one to four
* floating-point or integer numbers, as defined by the OpenGL specification.
* There are various places in OpenGL where the user can specify a compressed
* there are.
*/
#define MALI_NR_CHANNELS(n) ((n - 1) << 3)
+#define MALI_EXTRACT_CHANNELS(fmt) ((((fmt) >> 3) & 3) + 1)
/* If the high 3 bits are 3 to 6, then the low 3 bits say how big each
* component is, except the special MALI_CHANNEL_FLOAT which overrides what the
#define MALI_EXTRACT_BITS(fmt) (fmt & 0x7)
enum mali_format {
+ /* Not all formats are in fact available, need to query dynamically to
+ * check. Factory settings for Juno enables only ETC2 and ASTC, no
+ * DXT/RGTC formats.
+ * */
+
+ /* 0x0 invalid */
MALI_ETC2_RGB8 = MALI_FORMAT_COMPRESSED | 0x1,
MALI_ETC2_R11_UNORM = MALI_FORMAT_COMPRESSED | 0x2,
MALI_ETC2_RGBA8 = MALI_FORMAT_COMPRESSED | 0x3,
MALI_ETC2_RG11_UNORM = MALI_FORMAT_COMPRESSED | 0x4,
- MALI_ETC2_R11_SNORM = MALI_FORMAT_COMPRESSED | 0x11,
- MALI_ETC2_RG11_SNORM = MALI_FORMAT_COMPRESSED | 0x12,
+ /* 0x5 reserved */
+ MALI_NXR = MALI_FORMAT_COMPRESSED | 0x6, /* Nokia eXtended Range */
+ MALI_BC1_UNORM = MALI_FORMAT_COMPRESSED | 0x7, /* DXT1 */
+ MALI_BC2_UNORM = MALI_FORMAT_COMPRESSED | 0x8, /* DXT3 */
+ MALI_BC3_UNORM = MALI_FORMAT_COMPRESSED | 0x9, /* DXT5 */
+ MALI_BC4_UNORM = MALI_FORMAT_COMPRESSED | 0xA, /* RGTC1_UNORM */
+ MALI_BC4_SNORM = MALI_FORMAT_COMPRESSED | 0xB, /* RGTC1_SNORM */
+ MALI_BC5_UNORM = MALI_FORMAT_COMPRESSED | 0xC, /* RGTC2_UNORM */
+ MALI_BC5_SNORM = MALI_FORMAT_COMPRESSED | 0xD, /* RGTC2_SNORM */
+ MALI_BC6H_UF16 = MALI_FORMAT_COMPRESSED | 0xE,
+ MALI_BC6H_SF16 = MALI_FORMAT_COMPRESSED | 0xF,
+ MALI_BC7_UNORM = MALI_FORMAT_COMPRESSED | 0x10,
+ MALI_ETC2_R11_SNORM = MALI_FORMAT_COMPRESSED | 0x11, /* EAC_SNORM */
+ MALI_ETC2_RG11_SNORM = MALI_FORMAT_COMPRESSED | 0x12, /* EAC_SNORM */
MALI_ETC2_RGB8A1 = MALI_FORMAT_COMPRESSED | 0x13,
- MALI_ASTC_SRGB_SUPP = MALI_FORMAT_COMPRESSED | 0x16,
- MALI_ASTC_HDR_SUPP = MALI_FORMAT_COMPRESSED | 0x17,
+ MALI_ASTC_3D_LDR = MALI_FORMAT_COMPRESSED | 0x14,
+ MALI_ASTC_3D_HDR = MALI_FORMAT_COMPRESSED | 0x15,
+ MALI_ASTC_2D_LDR = MALI_FORMAT_COMPRESSED | 0x16,
+ MALI_ASTC_2D_HDR = MALI_FORMAT_COMPRESSED | 0x17,
MALI_RGB565 = MALI_FORMAT_SPECIAL | 0x0,
MALI_RGB5_X1_UNORM = MALI_FORMAT_SPECIAL | 0x1,
};
-/* Alpha coverage is encoded as 4-bits (from a clampf), with inversion
- * literally performing a bitwise invert. This function produces slightly wrong
- * results and I'm not sure why; some rounding issue I suppose... */
-
-#define MALI_ALPHA_COVERAGE(clampf) ((uint16_t) (int) (clampf * 15.0f))
-#define MALI_GET_ALPHA_COVERAGE(nibble) ((float) nibble / 15.0f)
-
/* Applies to midgard1.flags_lo */
/* Should be set when the fragment shader updates the depth value. */
* are side effects, hence it interacts with early-z. */
#define MALI_WRITES_GLOBAL (1 << 9)
-#define MALI_READS_TILEBUFFER (1 << 12)
+#define MALI_READS_TILEBUFFER (1 << 10)
/* Applies to midgard1.flags_hi */
u32 unknown2_2;
- u16 alpha_coverage;
+ /* Generated from SAMPLE_COVERAGE_VALUE and SAMPLE_COVERAGE_INVERT. See
+ * 13.8.3 ("Multisample Fragment Operations") in the OpenGL ES 3.2
+ * specification. Only matters when multisampling is enabled. */
+ u16 coverage_mask;
+
u16 unknown2_3;
u8 stencil_mask_front;
u8 stencil_mask_back;
u16 unknown2_4;
- struct mali_stencil_test stencil_front;
- struct mali_stencil_test stencil_back;
+ struct mali_stencil_packed stencil_front;
+ struct mali_stencil_packed stencil_back;
union {
struct {
/* ORed into an MFBD address to specify the fbx section is included */
#define MALI_MFBD_TAG_EXTRA (0x2)
-/* Uniform buffer objects are 64-bit fields divided as:
- *
- * u64 size : 10;
- * mali_ptr ptr : 64 - 10;
- *
- * The size is actually the size minus 1 (MALI_POSITIVE), in units of 16 bytes.
- * This gives a maximum of 2^14 bytes, which just so happens to be the GL
- * minimum-maximum for GL_MAX_UNIFORM_BLOCK_SIZE.
- *
- * The pointer is missing the bottom 2 bits and top 8 bits. The top 8 bits
- * should be 0 for userspace pointers, according to
- * https://lwn.net/Articles/718895/. By reusing these bits, we can make each
- * entry in the table only 64 bits.
- */
-
-#define MALI_MAKE_UBO(elements, ptr) \
- (MALI_POSITIVE((elements)) | (((ptr) >> 2) << 10))
-
/* On Bifrost, these fields are the same between the vertex and tiler payloads.
* They also seem to be the same between Bifrost and Midgard. They're shared in
* fused payloads.
#define MALI_POSITIVE(dim) (dim - 1)
-/* Used with wrapping. Unclear what top bit conveys */
-
-enum mali_wrap_mode {
- MALI_WRAP_REPEAT = 0x8 | 0x0,
- MALI_WRAP_CLAMP_TO_EDGE = 0x8 | 0x1,
- MALI_WRAP_CLAMP = 0x8 | 0x2,
- MALI_WRAP_CLAMP_TO_BORDER = 0x8 | 0x3,
- MALI_WRAP_MIRRORED_REPEAT = 0x8 | 0x4 | 0x0,
- MALI_WRAP_MIRRORED_CLAMP_TO_EDGE = 0x8 | 0x4 | 0x1,
- MALI_WRAP_MIRRORED_CLAMP = 0x8 | 0x4 | 0x2,
- MALI_WRAP_MIRRORED_CLAMP_TO_BORDER = 0x8 | 0x4 | 0x3,
-};
-
-/* Shared across both command stream and Midgard, and even with Bifrost */
-
-enum mali_texture_type {
- MALI_TEX_CUBE = 0x0,
- MALI_TEX_1D = 0x1,
- MALI_TEX_2D = 0x2,
- MALI_TEX_3D = 0x3
-};
-
/* 8192x8192 */
#define MAX_MIP_LEVELS (13)
/* For each pointer, there is an address and optionally also a stride */
#define MAX_ELEMENTS (2)
-/* It's not known why there are 4-bits allocated -- this enum is almost
- * certainly incomplete */
-
-enum mali_texture_layout {
- /* For a Z/S texture, this is linear */
- MALI_TEXTURE_TILED = 0x1,
-
- /* Z/S textures cannot be tiled */
- MALI_TEXTURE_LINEAR = 0x2,
-
- /* 16x16 sparse */
- MALI_TEXTURE_AFBC = 0xC
-};
-
-/* Corresponds to the type passed to glTexImage2D and so forth */
-
-struct mali_texture_format {
- unsigned swizzle : 12;
- enum mali_format format : 8;
-
- unsigned srgb : 1;
- unsigned unknown1 : 1;
-
- enum mali_texture_type type : 2;
- enum mali_texture_layout layout : 4;
-
- /* Always set */
- unsigned unknown2 : 1;
-
- /* Set to allow packing an explicit stride */
- unsigned manual_stride : 1;
-
- unsigned zero : 2;
-} __attribute__((packed));
-
-struct mali_texture_descriptor {
- uint16_t width;
- uint16_t height;
- uint16_t depth;
- uint16_t array_size;
-
- struct mali_texture_format format;
-
- uint16_t unknown3;
-
- /* One for non-mipmapped, zero for mipmapped */
- uint8_t unknown3A;
-
- /* Zero for non-mipmapped, (number of levels - 1) for mipmapped */
- uint8_t levels;
-
- /* Swizzling is a single 32-bit word, broken up here for convenience.
- * Here, swizzling refers to the ES 3.0 texture parameters for channel
- * level swizzling, not the internal pixel-level swizzling which is
- * below OpenGL's reach */
-
- unsigned swizzle : 12;
- unsigned swizzle_zero : 20;
-
- uint32_t unknown5;
- uint32_t unknown6;
- uint32_t unknown7;
-} __attribute__((packed));
-
-/* While Midgard texture descriptors are variable length, Bifrost descriptors
- * are fixed like samplers with more pointers to expand if necessary */
-
-struct bifrost_texture_descriptor {
- unsigned format_unk : 4; /* 2 */
- enum mali_texture_type type : 2;
- unsigned zero : 4;
- unsigned format_swizzle : 12;
- enum mali_format format : 8;
- unsigned srgb : 1;
- unsigned format_unk3 : 1; /* 0 */
-
- uint16_t width; /* MALI_POSITIVE */
- uint16_t height; /* MALI_POSITIVE */
-
- /* OpenGL swizzle */
- unsigned swizzle : 12;
- enum mali_texture_layout layout : 4;
- uint8_t levels : 8; /* Number of levels-1 if mipmapped, 0 if not */
- unsigned unk1 : 8;
-
- unsigned levels_unk : 24; /* 0 */
- unsigned level_2 : 8; /* Number of levels, again? */
-
- mali_ptr payload;
-
- uint16_t array_size;
- uint16_t unk4;
-
- uint16_t depth;
- uint16_t unk5;
-} __attribute__((packed));
-
-/* filter_mode */
-
-#define MALI_SAMP_MAG_NEAREST (1 << 0)
-#define MALI_SAMP_MIN_NEAREST (1 << 1)
-
-/* TODO: What do these bits mean individually? Only seen set together */
-
-#define MALI_SAMP_MIP_LINEAR_1 (1 << 3)
-#define MALI_SAMP_MIP_LINEAR_2 (1 << 4)
-
-/* Flag in filter_mode, corresponding to OpenCL's NORMALIZED_COORDS_TRUE
- * sampler_t flag. For typical OpenGL textures, this is always set. */
-
-#define MALI_SAMP_NORM_COORDS (1 << 5)
-
/* Used for lod encoding. Thanks @urjaman for pointing out these routines can
* be cleaned up a lot. */
return (int) (x * 256.0);
}
-struct mali_sampler_descriptor {
- uint16_t filter_mode;
-
- /* Fixed point, signed.
- * Upper 7 bits before the decimal point, although it caps [0-31].
- * Lower 8 bits after the decimal point: int(round(x * 256)) */
-
- int16_t lod_bias;
- int16_t min_lod;
- int16_t max_lod;
-
- /* All one word in reality, but packed a bit. Comparisons are flipped
- * from OpenGL. */
-
- enum mali_wrap_mode wrap_s : 4;
- enum mali_wrap_mode wrap_t : 4;
- enum mali_wrap_mode wrap_r : 4;
- enum mali_func compare_func : 3;
-
- /* No effect on 2D textures. For cubemaps, set for ES3 and clear for
- * ES2, controlling seamless cubemapping */
- unsigned seamless_cube_map : 1;
-
- unsigned zero : 16;
-
- uint32_t zero2;
- float border_color[4];
-} __attribute__((packed));
-
-/* Bifrost sampler descriptors look pretty similar */
-
-#define BIFROST_SAMP_MIN_NEAREST (1)
-#define BIFROST_SAMP_MAG_LINEAR (1)
-
-struct bifrost_sampler_descriptor {
- uint8_t unk1;
-
- enum mali_wrap_mode wrap_r : 4;
- enum mali_wrap_mode wrap_t : 4;
- enum mali_wrap_mode wrap_s : 4;
- uint8_t unk8 : 4;
-
- uint8_t unk2 : 1;
- uint8_t norm_coords : 1;
- uint8_t unk3 : 1;
- uint8_t min_filter : 1;
- uint8_t zero1 : 1;
- uint8_t mag_filter : 1;
- uint8_t mip_filter : 1;
-
- int16_t min_lod;
- int16_t max_lod;
-
- uint64_t zero2;
- uint64_t zero3;
- uint64_t zero4;
-} __attribute__((packed));
-
-/* viewport0/viewport1 form the arguments to glViewport. viewport1 is
- * modified by MALI_POSITIVE; viewport0 is as-is.
- */
-
-struct mali_viewport {
- /* XY clipping planes */
- float clip_minx;
- float clip_miny;
- float clip_maxx;
- float clip_maxy;
-
- /* Depth clipping planes */
- float clip_minz;
- float clip_maxz;
-
- u16 viewport0[2];
- u16 viewport1[2];
-} __attribute__((packed));
-
/* From presentations, 16x16 tiles externally. Use shift for fast computation
* of tile numbers. */
u32 weights[8];
};
-enum mali_block_format {
- MALI_BLOCK_TILED = 0x0,
- MALI_BLOCK_UNKNOWN = 0x1,
- MALI_BLOCK_LINEAR = 0x2,
- MALI_BLOCK_AFBC = 0x3,
-};
-
struct mali_sfbd_format {
/* 0x1 */
unsigned unk1 : 6;
/* More below this, maybe */
} __attribute__((packed));
-/* Format bits for the render target flags */
-#define MALI_MFBD_FORMAT_MSAA (1 << 1)
-#define MALI_MFBD_FORMAT_SRGB (1 << 2)
+#define MALI_MFBD_FORMAT_SRGB (1 << 0)
struct mali_rt_format {
unsigned unk1 : 32;
unsigned unk3 : 4;
unsigned unk4 : 1;
enum mali_block_format block : 2;
- unsigned flags : 4;
+ enum mali_msaa msaa : 2;
+ unsigned flags : 2;
unsigned swizzle : 12;
mali_ptr framebuffer;
u32 zero2 : 4;
- u32 framebuffer_stride : 28; // in units of bytes
- u32 zero3;
+ u32 framebuffer_stride : 28; // in units of bytes, row to next
+ u32 layer_stride; /* For multisample rendering */
u32 clear_color_1; // RGBA8888 from glClear, actually used by hardware
u32 clear_color_2; // always equal, but unclear function?
*/
/* flags_hi */
-#define MALI_EXTRA_PRESENT (0x10)
+#define MALI_EXTRA_PRESENT (0x1)
/* flags_lo */
#define MALI_EXTRA_ZS (0x4)
unsigned flags_lo : 4;
enum mali_block_format zs_block : 2;
- unsigned flags_hi : 26;
+
+ /* Number of samples in Z/S attachment, MALI_POSITIVE. So zero for
+ * 1-sample (non-MSAA), 0x3 for MSAA 4x, etc */
+ unsigned zs_samples : 4;
+ unsigned flags_hi : 22;
union {
/* Note: AFBC is only allowed for 24/8 combined depth/stencil. */
mali_ptr depth;
u32 depth_stride_zero : 4;
u32 depth_stride : 28;
- u32 zero1;
+ u32 depth_layer_stride;
mali_ptr stencil;
u32 stencil_stride_zero : 4;
u32 stencil_stride : 28;
- u32 zero2;
+ u32 stencil_layer_stride;
} ds_linear;
};
u32 zero3;
u16 width2, height2;
u32 unk1 : 19; // = 0x01000
- u32 rt_count_1 : 2; // off-by-one (use MALI_POSITIVE)
- u32 unk2 : 3; // = 0
+ u32 rt_count_1 : 3; // off-by-one (use MALI_POSITIVE)
+ u32 unk2 : 2; // = 0
u32 rt_count_2 : 3; // no off-by-one
u32 zero4 : 5;
/* 0x30 */