#define __PANFROST_JOB_H__
#include <stdint.h>
+#include <stdbool.h>
#include <panfrost-misc.h>
enum mali_job_type {
MALI_FUNC_ALWAYS = 7
};
-/* Same OpenGL, but mixed up. Why? Because forget me, that's why! */
-
-enum mali_alt_func {
- MALI_ALT_FUNC_NEVER = 0,
- MALI_ALT_FUNC_GREATER = 1,
- MALI_ALT_FUNC_EQUAL = 2,
- MALI_ALT_FUNC_GEQUAL = 3,
- MALI_ALT_FUNC_LESS = 4,
- MALI_ALT_FUNC_NOTEQUAL = 5,
- MALI_ALT_FUNC_LEQUAL = 6,
- MALI_ALT_FUNC_ALWAYS = 7
-};
-
/* Flags apply to unknown2_3? */
#define MALI_HAS_MSAA (1 << 0)
/* The top 3 bits specify how the bits of each component are interpreted. */
+/* e.g. ETC2_RGB8 */
+#define MALI_FORMAT_COMPRESSED (0 << 5)
+
/* e.g. R11F_G11F_B10F */
#define MALI_FORMAT_SPECIAL (2 << 5)
#define MALI_CHANNEL_FLOAT 7
enum mali_format {
+ 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,
+ MALI_ETC2_RGB8A1 = MALI_FORMAT_COMPRESSED | 0x13,
+ MALI_ASTC_SRGB_SUPP = MALI_FORMAT_COMPRESSED | 0x16,
+ MALI_ASTC_HDR_SUPP = MALI_FORMAT_COMPRESSED | 0x17,
+
MALI_RGB565 = MALI_FORMAT_SPECIAL | 0x0,
MALI_RGB5_A1_UNORM = MALI_FORMAT_SPECIAL | 0x2,
MALI_RGB10_A2_UNORM = MALI_FORMAT_SPECIAL | 0x3,
#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 */
+/* Applies to midgard1.flags_lo */
+
+/* Should be set when the fragment shader updates the depth value. */
+#define MALI_WRITES_Z (1 << 4)
/* Should the hardware perform early-Z testing? Normally should be set
* for performance reasons. Clear if you use: discard,
#define MALI_READS_ZS (1 << 8)
#define MALI_READS_TILEBUFFER (1 << 12)
+/* Applies to midgard1.flags_hi */
+
+/* Should be set when the fragment shader updates the stencil value. */
+#define MALI_WRITES_S (1 << 2)
+
/* The raw Midgard blend payload can either be an equation or a shader
* address, depending on the context */
} bifrost1;
struct {
unsigned uniform_buffer_count : 4;
- unsigned flags : 12;
+ unsigned flags_lo : 12;
- /* Whole number of uniform registers used, times two;
- * whole number of work registers used (no scale).
- */
+ /* vec4 units */
unsigned work_count : 5;
unsigned uniform_count : 5;
- unsigned unknown2 : 6;
+ unsigned flags_hi : 6;
} midgard1;
};
u16 job_index;
u16 job_dependency_index_1;
u16 job_dependency_index_2;
-
- union {
- u64 next_job_64;
- u32 next_job_32;
- };
+ u64 next_job;
} __attribute__((packed));
/* These concern exception_status */
u64 immediate;
} __attribute__((packed));
-/* Special attributes have a fixed index */
-#define MALI_SPECIAL_ATTRIBUTE_BASE 16
-#define MALI_VERTEX_ID (MALI_SPECIAL_ATTRIBUTE_BASE + 0)
-#define MALI_INSTANCE_ID (MALI_SPECIAL_ATTRIBUTE_BASE + 1)
-
/*
* Mali Attributes
*
* let shift=extra_flags=0. Stride is set to the image format's bytes-per-pixel
* (*NOT the row stride*). Size is set to the size of the image itself.
*
- * Special internal varyings (including gl_FrontFacing) could be seen as
- * IMAGE/INTERNAL as well as LINEAR, setting all fields set to zero and using a
- * special elements pseudo-pointer.
+ * Special internal attribtues and varyings (gl_VertexID, gl_FrontFacing, etc)
+ * use particular fixed addresses with modified structures.
*/
enum mali_attr_mode {
MALI_ATTR_MODULO = 3,
MALI_ATTR_NPOT_DIVIDE = 4,
MALI_ATTR_IMAGE = 5,
- MALI_ATTR_INTERNAL = 6
};
-/* Pseudo-address for gl_FrontFacing, used with INTERNAL. Same addres is used
- * for gl_FragCoord with IMAGE, needing a coordinate flip. Who knows. */
+/* Pseudo-address for gl_VertexID, gl_FragCoord, gl_FrontFacing */
+#define MALI_ATTR_VERTEXID (0x22)
+#define MALI_ATTR_INSTANCEID (0x24)
#define MALI_VARYING_FRAG_COORD (0x25)
#define MALI_VARYING_FRONT_FACING (0x26)
/* Used for comparison to check if an address is special. Mostly a guess, but
* it doesn't really matter. */
-#define MALI_VARYING_SPECIAL (0x100)
+#define MALI_RECORD_SPECIAL (0x100)
union mali_attr {
/* This is used for actual attributes. */
int32_t src_offset;
} __attribute__((packed));
-enum mali_fbd_type {
- MALI_SFBD = 0,
- MALI_MFBD = 1,
-};
-
-#define FBD_TYPE (1)
#define FBD_MASK (~0x3f)
+/* MFBD, rather than SFBD */
+#define MALI_MFBD (0x1)
+
/* ORed into an MFBD address to specify the fbx section is included */
#define MALI_MFBD_TAG_EXTRA (0x2)
#define MALI_DRAW_INDEXED_SHIFT (4)
#define MALI_DRAW_VARYING_SIZE (0x100)
+
+/* Set to use first vertex as the provoking vertex for flatshading. Clear to
+ * use the last vertex. This is the default in DX and VK, but not in GL. */
+
+#define MALI_DRAW_FLATSHADE_FIRST (0x800)
+
#define MALI_DRAW_PRIMITIVE_RESTART_FIXED_INDEX (0x10000)
struct mali_vertex_tiler_prefix {
*/
u32 invocation_count;
- u32 size_y_shift : 5;
- u32 size_z_shift : 5;
- u32 workgroups_x_shift : 6;
- u32 workgroups_y_shift : 6;
- u32 workgroups_z_shift : 6;
- /* This is max(workgroups_x_shift, 2) in all the cases I've seen. */
- u32 workgroups_x_shift_2 : 4;
+ /* Bitfield for shifts:
+ *
+ * size_y_shift : 5
+ * size_z_shift : 5
+ * workgroups_x_shift : 6
+ * workgroups_y_shift : 6
+ * workgroups_z_shift : 6
+ * workgroups_x_shift_2 : 4
+ */
+ u32 invocation_shifts;
u32 draw_mode : 4;
u32 unknown_draw : 22;
u64 zero8;
} __attribute__((packed));
-struct bifrost_scratchpad {
- u32 zero;
- u32 flags; // = 0x1f
- /* This is a pointer to a CPU-inaccessible buffer, 16 pages, allocated
- * during startup. It seems to serve the same purpose as the
- * gpu_scratchpad in the SFBD for Midgard, although it's slightly
- * larger.
- */
- mali_ptr gpu_scratchpad;
-} __attribute__((packed));
-
struct mali_vertex_tiler_postfix {
/* Zero for vertex jobs. Pointer to the position (gl_Position) varying
* output from the vertex shader for tiler jobs.
u64 viewport;
u64 occlusion_counter; /* A single bit as far as I can tell */
- /* Note: on Bifrost, this isn't actually the FBD. It points to
- * bifrost_scratchpad instead. However, it does point to the same thing
- * in vertex and tiler jobs.
- */
- mali_ptr framebuffer;
+ /* On Bifrost, this points directly to a mali_shared_memory structure.
+ * On Midgard, this points to a framebuffer (either SFBD or MFBD as
+ * tagged), which embeds a mali_shared_memory structure */
+ mali_ptr shared_memory;
} __attribute__((packed));
struct midgard_payload_vertex_tiler {
#define MALI_POSITIVE(dim) (dim - 1)
-/* Opposite of MALI_POSITIVE, found in the depth_units field */
-
-#define MALI_NEGATIVE(dim) (dim + 1)
-
-/* Used with wrapping. Incomplete (this is a 4-bit field...) */
+/* Used with wrapping. Unclear what top bit conveys */
enum mali_wrap_mode {
- MALI_WRAP_REPEAT = 0x8,
- MALI_WRAP_CLAMP_TO_EDGE = 0x9,
- MALI_WRAP_CLAMP_TO_BORDER = 0xB,
- MALI_WRAP_MIRRORED_REPEAT = 0xC
+ 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 */
uint32_t unknown5;
uint32_t unknown6;
uint32_t unknown7;
-
- mali_ptr payload[MAX_MIP_LEVELS * MAX_CUBE_FACES * MAX_ELEMENTS];
} __attribute__((packed));
/* filter_mode */
#define DECODE_FIXED_16(x) ((float) (x / 256.0))
-static inline uint16_t
-FIXED_16(float x)
+static inline int16_t
+FIXED_16(float x, bool allow_negative)
{
/* Clamp inputs, accounting for float error */
float max_lod = (32.0 - (1.0 / 512.0));
+ float min_lod = allow_negative ? -max_lod : 0.0;
- x = ((x > max_lod) ? max_lod : ((x < 0.0) ? 0.0 : x));
+ x = ((x > max_lod) ? max_lod : ((x < min_lod) ? min_lod : x));
return (int) (x * 256.0);
}
struct mali_sampler_descriptor {
uint16_t filter_mode;
- /* Fixed point. Upper 8-bits is before the decimal point, although it
- * caps [0-31]. Lower 8-bits is after the decimal point: int(round(x *
- * 256)) */
+ /* 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)) */
- uint16_t lod_bias;
- uint16_t min_lod;
- uint16_t max_lod;
+ int16_t lod_bias;
+ int16_t min_lod;
+ int16_t max_lod;
- /* All one word in reality, but packed a bit */
+ /* 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_alt_func compare_func : 3;
+ enum mali_func compare_func : 3;
/* No effect on 2D textures. For cubemaps, set for ES3 and clear for
* ES2, controlling seamless cubemapping */
unsigned unk3 : 4;
};
-struct mali_single_framebuffer {
- u32 unknown1;
- u32 unknown2;
- u64 unknown_address_0;
- u64 zero1;
- u64 zero0;
+/* Shared structure at the start of framebuffer descriptors, or used bare for
+ * compute jobs, configuring stack and shared memory */
+
+struct mali_shared_memory {
+ u32 stack_shift : 4;
+ u32 unk0 : 28;
+
+ /* Configuration for shared memory for compute shaders.
+ * shared_workgroup_count is logarithmic and may be computed for a
+ * compute shader using shared memory as:
+ *
+ * shared_workgroup_count = MAX2(ceil(log2(count_x)) + ... + ceil(log2(count_z), 10)
+ *
+ * For compute shaders that don't use shared memory, or non-compute
+ * shaders, this is set to ~0
+ */
+
+ u32 shared_workgroup_count : 5;
+ u32 shared_unk1 : 3;
+ u32 shared_shift : 4;
+ u32 shared_zero : 20;
+
+ mali_ptr scratchpad;
+ /* For compute shaders, the RAM backing of workgroup-shared memory. For
+ * fragment shaders on Bifrost, apparently multisampling locations */
+
+ mali_ptr shared_memory;
+ mali_ptr unknown1;
+} __attribute__((packed));
+
+
+
+struct mali_single_framebuffer {
+ struct mali_shared_memory shared_memory;
struct mali_sfbd_format format;
u32 clear_flags;
/* More below this, maybe */
} __attribute__((packed));
-/* On Midgard, this "framebuffer descriptor" is used for the framebuffer field
- * of compute jobs. Superficially resembles a single framebuffer descriptor */
-
-struct mali_compute_fbd {
- u32 unknown1[8];
-} __attribute__((packed));
-
/* Format bits for the render target flags */
#define MALI_MFBD_FORMAT_MSAA (1 << 1)
* - TODO: Anything else?
*/
-/* Flags field: note, these are guesses */
+/* flags_hi */
+#define MALI_EXTRA_PRESENT (0x10)
-#define MALI_EXTRA_PRESENT (0x400)
-#define MALI_EXTRA_AFBC (0x20)
-#define MALI_EXTRA_AFBC_ZS (0x10)
+/* flags_lo */
#define MALI_EXTRA_ZS (0x4)
struct bifrost_fb_extra {
/* Each tile has an 8 byte checksum, so the stride is "width in tiles * 8" */
u32 checksum_stride;
- u32 flags;
+ unsigned flags_lo : 4;
+ enum mali_block_format zs_block : 2;
+ unsigned flags_hi : 26;
union {
/* Note: AFBC is only allowed for 24/8 combined depth/stencil. */
#define MALI_MFBD_EXTRA (1 << 13)
struct bifrost_framebuffer {
- u32 unk0; // = 0x10
+ struct mali_shared_memory shared_memory;
- u32 unknown2; // = 0x1f, same as SFBD
- mali_ptr scratchpad;
-
- /* 0x10 */
- mali_ptr sample_locations;
- mali_ptr unknown1;
/* 0x20 */
u16 width1, height1;
u32 zero3;