#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,
-};
-
-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 */
-};
-
-/* Applies to tiler_gl_enables */
-
-#define MALI_OCCLUSION_QUERY (1 << 3)
-#define MALI_OCCLUSION_PRECISE (1 << 4)
-
-/* Set for a glFrontFace(GL_CCW) in a Y=0=TOP coordinate system (like Gallium).
- * In OpenGL, this would corresponds to glFrontFace(GL_CW). Mesa and the blob
- * disagree about how to do viewport flipping, so the blob actually sets this
- * for GL_CW but then has a negative viewport stride */
-
-#define MALI_FRONT_CCW_TOP (1 << 5)
-
-#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)
-#define MALI_CAN_DISCARD (1 << 5)
-
-/* Applies on SFBD systems, specifying that programmable blending is in use */
-#define MALI_HAS_BLEND_SHADER (1 << 6)
-
-/* func is mali_func */
-#define MALI_DEPTH_FUNC(func) (func << 8)
-#define MALI_GET_DEPTH_FUNC(flags) ((flags >> 8) & 0x7)
-#define MALI_DEPTH_FUNC_MASK MALI_DEPTH_FUNC(0x7)
-
-#define MALI_DEPTH_WRITEMASK (1 << 11)
+#include <inttypes.h>
-/* 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_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)
-#define MALI_MASK_A (1 << 3)
+typedef uint8_t u8;
+typedef uint16_t u16;
+typedef uint32_t u32;
+typedef uint64_t u64;
+typedef uint64_t mali_ptr;
enum mali_nondominant_mode {
MALI_BLEND_NON_MIRROR = 0,
unsigned complement_dominant : 1;
} __attribute__((packed));
-struct mali_blend_equation {
- /* Of type mali_blend_mode */
- unsigned rgb_mode : 12;
- unsigned alpha_mode : 12;
-
- unsigned zero1 : 4;
-
- /* Corresponds to MALI_MASK_* above and glColorMask arguments */
-
- unsigned color_mask : 4;
-} __attribute__((packed));
-
-/* Used with channel swizzling */
-enum mali_channel {
- MALI_CHANNEL_RED = 0,
- MALI_CHANNEL_GREEN = 1,
- MALI_CHANNEL_BLUE = 2,
- MALI_CHANNEL_ALPHA = 3,
- MALI_CHANNEL_ZERO = 4,
- MALI_CHANNEL_ONE = 5,
- MALI_CHANNEL_RESERVED_0 = 6,
- 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
#define MALI_CHANNEL_FLOAT 7
#define MALI_EXTRACT_BITS(fmt) (fmt & 0x7)
-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_X1_UNORM = MALI_FORMAT_SPECIAL | 0x1,
- MALI_RGB5_A1_UNORM = MALI_FORMAT_SPECIAL | 0x2,
- MALI_RGB10_A2_UNORM = MALI_FORMAT_SPECIAL | 0x3,
- MALI_RGB10_A2_SNORM = MALI_FORMAT_SPECIAL | 0x5,
- MALI_RGB10_A2UI = MALI_FORMAT_SPECIAL | 0x7,
- MALI_RGB10_A2I = MALI_FORMAT_SPECIAL | 0x9,
-
- MALI_RGB332_UNORM = MALI_FORMAT_SPECIAL | 0xb,
- MALI_RGB233_UNORM = MALI_FORMAT_SPECIAL | 0xc,
-
- MALI_Z24X8_UNORM = MALI_FORMAT_SPECIAL | 0xd,
- MALI_R32_FIXED = MALI_FORMAT_SPECIAL | 0x11,
- MALI_RG32_FIXED = MALI_FORMAT_SPECIAL | 0x12,
- MALI_RGB32_FIXED = MALI_FORMAT_SPECIAL | 0x13,
- MALI_RGBA32_FIXED = MALI_FORMAT_SPECIAL | 0x14,
- MALI_R11F_G11F_B10F = MALI_FORMAT_SPECIAL | 0x19,
- MALI_R9F_G9F_B9F_E5F = MALI_FORMAT_SPECIAL | 0x1b,
- /* Only used for varyings, to indicate the transformed gl_Position */
- MALI_VARYING_POS = MALI_FORMAT_SPECIAL | 0x1e,
- /* Only used for varyings, to indicate that the write should be
- * discarded.
- */
- MALI_VARYING_DISCARD = MALI_FORMAT_SPECIAL | 0x1f,
-
- MALI_R8_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_8,
- MALI_R16_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_16,
- MALI_R32_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_32,
- MALI_RG8_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_8,
- MALI_RG16_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_16,
- MALI_RG32_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_32,
- MALI_RGB8_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_8,
- MALI_RGB16_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_16,
- MALI_RGB32_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_32,
- MALI_RGBA8_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_8,
- MALI_RGBA16_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_16,
- MALI_RGBA32_SNORM = MALI_FORMAT_SNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_32,
-
- MALI_R8UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_8,
- MALI_R16UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_16,
- MALI_R32UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_32,
- MALI_RG8UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_8,
- MALI_RG16UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_16,
- MALI_RG32UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_32,
- MALI_RGB8UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_8,
- MALI_RGB16UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_16,
- MALI_RGB32UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_32,
- MALI_RGBA8UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_8,
- MALI_RGBA16UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_16,
- MALI_RGBA32UI = MALI_FORMAT_UINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_32,
-
- MALI_R8_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_8,
- MALI_R16_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_16,
- MALI_R32_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_32,
- MALI_R32F = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(1) | MALI_CHANNEL_FLOAT,
- MALI_RG8_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_8,
- MALI_RG16_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_16,
- MALI_RG32_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_32,
- MALI_RG32F = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(2) | MALI_CHANNEL_FLOAT,
- MALI_RGB8_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_8,
- MALI_RGB16_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_16,
- MALI_RGB32_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_32,
- MALI_RGB32F = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(3) | MALI_CHANNEL_FLOAT,
- MALI_RGBA4_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_4,
- MALI_RGBA8_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_8,
- MALI_RGBA16_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_16,
- MALI_RGBA32_UNORM = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_32,
- MALI_RGBA32F = MALI_FORMAT_UNORM | MALI_NR_CHANNELS(4) | MALI_CHANNEL_FLOAT,
-
- MALI_R8I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_8,
- MALI_R16I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_16,
- MALI_R32I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_32,
- MALI_R16F = MALI_FORMAT_SINT | MALI_NR_CHANNELS(1) | MALI_CHANNEL_FLOAT,
- MALI_RG8I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_8,
- MALI_RG16I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_16,
- MALI_RG32I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_32,
- MALI_RG16F = MALI_FORMAT_SINT | MALI_NR_CHANNELS(2) | MALI_CHANNEL_FLOAT,
- MALI_RGB8I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_8,
- MALI_RGB16I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_16,
- MALI_RGB32I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_32,
- MALI_RGB16F = MALI_FORMAT_SINT | MALI_NR_CHANNELS(3) | MALI_CHANNEL_FLOAT,
- MALI_RGBA8I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_8,
- MALI_RGBA16I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_16,
- MALI_RGBA32I = MALI_FORMAT_SINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_32,
- MALI_RGBA16F = MALI_FORMAT_SINT | MALI_NR_CHANNELS(4) | MALI_CHANNEL_FLOAT,
-
- MALI_RGBA4 = MALI_FORMAT_SPECIAL2 | 0x8,
- MALI_RGBA8_2 = MALI_FORMAT_SPECIAL2 | 0xd,
- MALI_RGB10_A2_2 = MALI_FORMAT_SPECIAL2 | 0xe,
-};
-
-
-/* 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. */
-#define MALI_WRITES_Z (1 << 4)
-
-/* Should the hardware perform early-Z testing? Set if the shader does not use
- * discard, alpha-to-coverage, shader depth writes, and if the shader has no
- * side effects (writes to global memory or images) unless early-z testing is
- * forced in the shader.
- */
-
-#define MALI_EARLY_Z (1 << 6)
-
-/* Should the hardware calculate derivatives (via helper invocations)? Set in a
- * fragment shader that uses texturing or derivative functions */
-
-#define MALI_HELPER_INVOCATIONS (1 << 7)
-
-/* Flags denoting the fragment shader's use of tilebuffer readback. If the
- * shader might read any part of the tilebuffer, set MALI_READS_TILEBUFFER. If
- * it might read depth/stencil in particular, also set MALI_READS_ZS */
-
-#define MALI_READS_ZS (1 << 8)
-
-/* The shader might write to global memory (via OpenCL, SSBOs, or images).
- * Reading is okay, as are ordinary writes to the tilebuffer/varyings. Setting
- * incurs a performance penalty. On a fragment shader, this bit implies there
- * are side effects, hence it interacts with early-z. */
-#define MALI_WRITES_GLOBAL (1 << 9)
-
-#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)
-
-/* Mode to suppress generation of Infinity and NaN values by clamping inf
- * (-inf) to MAX_FLOAT (-MIN_FLOAT) and flushing NaN to 0.0
- *
- * Compare suppress_inf/suppress_nan flags on the Bifrost clause header for the
- * same functionality.
- *
- * This is not conformant on GLES3 or OpenCL, but is optional on GLES2, where
- * it works around app bugs (e.g. in glmark2-es2 -bterrain with FP16).
- */
-#define MALI_SUPPRESS_INF_NAN (1 << 3)
-
-/* Flags for bifrost1.unk1 */
-
-/* Shader uses less than 32 registers, partitioned as [R0, R15] U [R48, R63],
- * allowing for full thread count. If clear, the full [R0, R63] register set is
- * available at half thread count */
-#define MALI_BIFROST_FULL_THREAD (1 << 9)
-
-/* Enable early-z testing (presumably). This flag may not be set if the shader:
- *
- * - Uses blending
- * - Uses discard
- * - Writes gl_FragDepth
- *
- * This differs from Midgard which sets the MALI_EARLY_Z flag even with
- * blending, although I've begun to suspect that flag does not in fact enable
- * EARLY_Z alone. */
-#define MALI_BIFROST_EARLY_Z (1 << 15)
-
-/* First clause type is ATEST */
-#define MALI_BIFROST_FIRST_ATEST (1 << 26)
-
/* The raw Midgard blend payload can either be an equation or a shader
* address, depending on the context */
mali_ptr shader;
struct {
- struct mali_blend_equation equation;
+ struct mali_blend_equation_packed equation;
float constant;
};
};
-/* We need to load the tilebuffer to blend (i.e. the destination factor is not
- * ZERO) */
-
-#define MALI_BLEND_LOAD_TIB (0x1)
-
-/* A blend shader is used to blend this render target */
-#define MALI_BLEND_MRT_SHADER (0x2)
-
-/* On MRT Midgard systems (using an MFBD), each render target gets its own
- * blend descriptor */
-
-#define MALI_BLEND_SRGB (0x400)
-
-/* Dithering is specified here for MFBD, otherwise NO_DITHER for SFBD */
-#define MALI_BLEND_NO_DITHER (0x800)
-
struct midgard_blend_rt {
- /* Flags base value of 0x200 to enable the render target.
- * OR with 0x1 for blending (anything other than REPLACE).
- * OR with 0x2 for programmable blending
- * OR with MALI_BLEND_SRGB for implicit sRGB
- */
-
- u64 flags;
+ struct mali_blend_flags_packed flags;
+ u32 zero;
union midgard_blend blend;
} __attribute__((packed));
*/
u16 constant;
- struct mali_blend_equation equation;
+ struct mali_blend_equation_packed equation;
/*
* - 0x19 normally
};
} __attribute__((packed));
-/* Descriptor for the shader. Following this is at least one, up to four blend
- * descriptors for each active render target */
-
-struct mali_shader_meta {
- mali_ptr shader;
- u16 sampler_count;
- u16 texture_count;
- u16 attribute_count;
- u16 varying_count;
-
- union {
- struct {
- u32 uniform_buffer_count : 4;
- u32 unk1 : 28; // = 0x800000 for vertex, 0x958020 for tiler
- } bifrost1;
- struct {
- unsigned uniform_buffer_count : 4;
- unsigned flags_lo : 12;
-
- /* vec4 units */
- unsigned work_count : 5;
- unsigned uniform_count : 5;
- unsigned flags_hi : 6;
- } midgard1;
- };
-
- /* Same as glPolygoOffset() arguments */
- float depth_units;
- float depth_factor;
-
- u32 unknown2_2;
-
- u16 alpha_coverage;
- 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;
-
- union {
- struct {
- u32 unk3 : 7;
- /* On Bifrost, some system values are preloaded in
- * registers R55-R62 by the thread dispatcher prior to
- * the start of shader execution. This is a bitfield
- * with one entry for each register saying which
- * registers need to be preloaded. Right now, the known
- * values are:
- *
- * Vertex/compute:
- * - R55 : gl_LocalInvocationID.xy
- * - R56 : gl_LocalInvocationID.z + unknown in high 16 bits
- * - R57 : gl_WorkGroupID.x
- * - R58 : gl_WorkGroupID.y
- * - R59 : gl_WorkGroupID.z
- * - R60 : gl_GlobalInvocationID.x
- * - R61 : gl_GlobalInvocationID.y/gl_VertexID (without base)
- * - R62 : gl_GlobalInvocationID.z/gl_InstanceID (without base)
- *
- * Fragment:
- * - R55 : unknown, never seen (but the bit for this is
- * always set?)
- * - R56 : unknown (bit always unset)
- * - R57 : gl_PrimitiveID
- * - R58 : gl_FrontFacing in low bit, potentially other stuff
- * - R59 : u16 fragment coordinates (used to compute
- * gl_FragCoord.xy, together with sample positions)
- * - R60 : gl_SampleMask (used in epilog, so pretty
- * much always used, but the bit is always 0 -- is
- * this just always pushed?)
- * - R61 : gl_SampleMaskIn and gl_SampleID, used by
- * varying interpolation.
- * - R62 : unknown (bit always unset).
- *
- * Later GPUs (starting with Mali-G52?) support
- * preloading float varyings into r0-r7. This is
- * indicated by setting 0x40. There is no distinction
- * here between 1 varying and 2.
- */
- u32 preload_regs : 8;
- /* In units of 8 bytes or 64 bits, since the
- * uniform/const port loads 64 bits at a time.
- */
- u32 uniform_count : 7;
- u32 unk4 : 10; // = 2
- } bifrost2;
- struct {
- u32 unknown2_7;
- } midgard2;
- };
-
- u32 padding;
-
- /* Blending information for the older non-MRT Midgard HW. Check for
- * MALI_HAS_BLEND_SHADER to decide how to interpret.
- */
-
- union midgard_blend blend;
-} __attribute__((packed));
-
-/* This only concerns hardware jobs */
-
/* Possible values for job_descriptor_size */
#define MALI_JOB_32 0
u64 next_job;
} __attribute__((packed));
-/* These concern exception_status */
-
-/* Access type causing a fault, paralleling AS_FAULTSTATUS_* entries in the
- * kernel */
-
-enum mali_exception_access {
- /* Atomic in the kernel for MMU, but that doesn't make sense for a job
- * fault so it's just unused */
- MALI_EXCEPTION_ACCESS_NONE = 0,
-
- MALI_EXCEPTION_ACCESS_EXECUTE = 1,
- MALI_EXCEPTION_ACCESS_READ = 2,
- MALI_EXCEPTION_ACCESS_WRITE = 3
-};
-
/* Details about write_value from panfrost igt tests which use it as a generic
* dword write primitive */
* 3. If e <= 2^shift, then we need to use the round-down algorithm. Set
* magic_divisor = m - 1 and extra_flags = 1.
* 4. Otherwise, set magic_divisor = m and extra_flags = 0.
- *
- * Unrelated to instancing/actual attributes, images (the OpenCL kind) are
- * implemented as special attributes, denoted by MALI_ATTR_IMAGE. For images,
- * 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 attribtues and varyings (gl_VertexID, gl_FrontFacing, etc)
- * use particular fixed addresses with modified structures.
*/
-enum mali_attr_mode {
- MALI_ATTR_UNUSED = 0,
- MALI_ATTR_LINEAR = 1,
- MALI_ATTR_POT_DIVIDE = 2,
- MALI_ATTR_MODULO = 3,
- MALI_ATTR_NPOT_DIVIDE = 4,
- MALI_ATTR_IMAGE = 5,
-};
-
-/* 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)
-
-/* This magic "pseudo-address" is used as `elements` to implement
- * gl_PointCoord. When read from a fragment shader, it generates a point
- * coordinate per the OpenGL ES 2.0 specification. Flipped coordinate spaces
- * require an affine transformation in the shader. */
-
-#define MALI_VARYING_POINT_COORD (0x61)
-
-/* Used for comparison to check if an address is special. Mostly a guess, but
- * it doesn't really matter. */
-
-#define MALI_RECORD_SPECIAL (0x100)
-
-union mali_attr {
- /* This is used for actual attributes. */
- struct {
- /* The bottom 3 bits are the mode */
- mali_ptr elements : 64 - 8;
- u32 shift : 5;
- u32 extra_flags : 3;
- u32 stride;
- u32 size;
- };
- /* The entry after an NPOT_DIVIDE entry has this format. It stores
- * extra information that wouldn't fit in a normal entry.
- */
- struct {
- u32 unk; /* = 0x20 */
- u32 magic_divisor;
- u32 zero;
- /* This is the original, GL-level divisor. */
- u32 divisor;
- };
-} __attribute__((packed));
-
-struct mali_attr_meta {
- /* Vertex buffer index */
- u8 index;
-
- unsigned unknown1 : 2;
- unsigned swizzle : 12;
- enum mali_format format : 8;
-
- /* Always observed to be zero at the moment */
- unsigned unknown3 : 2;
-
- /* When packing multiple attributes in a buffer, offset addresses by
- * this value. Obscurely, this is signed. */
- int32_t src_offset;
-} __attribute__((packed));
-
#define FBD_MASK (~0x3f)
/* MFBD, rather than SFBD */
/* 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.
*/
-/* Applies to unknown_draw */
-
-#define MALI_DRAW_INDEXED_UINT8 (0x10)
-#define MALI_DRAW_INDEXED_UINT16 (0x20)
-#define MALI_DRAW_INDEXED_UINT32 (0x30)
-#define MALI_DRAW_INDEXED_SIZE (0x30)
-#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 {
- /* This is a dynamic bitfield containing the following things in this order:
- *
- * - gl_WorkGroupSize.x
- * - gl_WorkGroupSize.y
- * - gl_WorkGroupSize.z
- * - gl_NumWorkGroups.x
- * - gl_NumWorkGroups.y
- * - gl_NumWorkGroups.z
- *
- * The number of bits allocated for each number is based on the *_shift
- * fields below. For example, workgroups_y_shift gives the bit that
- * gl_NumWorkGroups.y starts at, and workgroups_z_shift gives the bit
- * that gl_NumWorkGroups.z starts at (and therefore one after the bit
- * that gl_NumWorkGroups.y ends at). The actual value for each gl_*
- * value is one more than the stored value, since if any of the values
- * are zero, then there would be no invocations (and hence no job). If
- * there were 0 bits allocated to a given field, then it must be zero,
- * and hence the real value is one.
- *
- * Vertex jobs reuse the same job dispatch mechanism as compute jobs,
- * effectively doing glDispatchCompute(1, vertex_count, instance_count)
- * where vertex count is the number of vertices.
- */
- u32 invocation_count;
-
- /* 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;
-
- /* This is the the same as workgroups_x_shift_2 in compute shaders, but
- * always 5 for vertex jobs and 6 for tiler jobs. I suspect this has
- * something to do with how many quads get put in the same execution
- * engine, which is a balance (you don't want to starve the engine, but
- * you also want to distribute work evenly).
- */
- u32 workgroups_x_shift_3 : 6;
-
-
- /* Negative of min_index. This is used to compute
- * the unbiased index in tiler/fragment shader runs.
- *
- * The hardware adds offset_bias_correction in each run,
- * so that absent an index bias, the first vertex processed is
- * genuinely the first vertex (0). But with an index bias,
- * the first vertex process is numbered the same as the bias.
- *
- * To represent this more conviniently:
- * unbiased_index = lower_bound_index +
- * index_bias +
- * offset_bias_correction
- *
- * This is done since the hardware doesn't accept a index_bias
- * and this allows it to recover the unbiased index.
- */
- int32_t offset_bias_correction;
- u32 zero1;
-
- /* Like many other strictly nonzero quantities, index_count is
- * subtracted by one. For an indexed cube, this is equal to 35 = 6
- * faces * 2 triangles/per face * 3 vertices/per triangle - 1. That is,
- * for an indexed draw, index_count is the number of actual vertices
- * rendered whereas invocation_count is the number of unique vertices
- * rendered (the number of times the vertex shader must be invoked).
- * For non-indexed draws, this is just equal to invocation_count. */
-
- u32 index_count;
-
- /* No hidden structure; literally just a pointer to an array of uint
- * indices (width depends on flags). Thanks, guys, for not making my
- * life insane for once! NULL for non-indexed draws. */
-
- u64 indices;
+ struct mali_invocation_packed invocation;
+ struct mali_primitive_packed primitive;
} __attribute__((packed));
/* Point size / line width can either be specified as a 32-bit float (for
u64 zeros[20];
} __attribute__((packed));
-struct bifrost_tiler_only {
- /* 0x20 */
- union midgard_primitive_size primitive_size;
-
- mali_ptr tiler_meta;
-
- u64 zero1, zero2, zero3, zero4, zero5, zero6;
-} __attribute__((packed));
-
-struct mali_vertex_tiler_postfix {
- u16 gl_enables; // 0x6 on Midgard, 0x2 on Bifrost
-
- /* Both zero for non-instanced draws. For instanced draws, a
- * decomposition of padded_num_vertices. See the comments about the
- * corresponding fields in mali_attr for context. */
-
- unsigned instance_shift : 5;
- unsigned instance_odd : 3;
-
- u8 zero4;
-
- /* Offset for first vertex in buffer */
- u32 offset_start;
-
- u64 zero5;
-
- /* Zero for vertex jobs. Pointer to the position (gl_Position) varying
- * output from the vertex shader for tiler jobs.
- */
-
- u64 position_varying;
-
- /* An array of mali_uniform_buffer_meta's. The size is given by the
- * shader_meta.
- */
- u64 uniform_buffers;
-
- /* On Bifrost, this is a pointer to an array of bifrost_texture_descriptor.
- * On Midgard, this is a pointer to an array of pointers to the texture
- * descriptors, number of pointers bounded by number of textures. The
- * indirection is needed to accomodate varying numbers and sizes of
- * texture descriptors */
- u64 textures;
-
- /* For OpenGL, from what I've seen, this is intimately connected to
- * texture_meta. cwabbott says this is not the case under Vulkan, hence
- * why this field is seperate (Midgard is Vulkan capable). Pointer to
- * array of sampler descriptors (which are uniform in size) */
- u64 sampler_descriptor;
-
- u64 uniforms;
- u64 shader;
- u64 attributes; /* struct attribute_buffer[] */
- u64 attribute_meta; /* attribute_meta[] */
- u64 varyings; /* struct attr */
- u64 varying_meta; /* pointer */
- u64 viewport;
- u64 occlusion_counter; /* A single bit as far as I can tell */
-
- /* 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 {
struct mali_vertex_tiler_prefix prefix;
- struct mali_vertex_tiler_postfix postfix;
-
+ struct mali_draw_packed postfix;
union midgard_primitive_size primitive_size;
} __attribute__((packed));
struct bifrost_payload_vertex {
struct mali_vertex_tiler_prefix prefix;
- struct mali_vertex_tiler_postfix postfix;
+ struct mali_draw_packed postfix;
} __attribute__((packed));
struct bifrost_payload_tiler {
struct mali_vertex_tiler_prefix prefix;
- struct bifrost_tiler_only tiler;
- struct mali_vertex_tiler_postfix postfix;
-} __attribute__((packed));
-
-struct bifrost_payload_fused {
- struct mali_vertex_tiler_prefix prefix;
- struct bifrost_tiler_only tiler;
- struct mali_vertex_tiler_postfix tiler_postfix;
- u64 padding; /* zero */
- struct mali_vertex_tiler_postfix vertex_postfix;
+ union midgard_primitive_size primitive_size;
+ mali_ptr tiler_meta;
+ u64 zero1, zero2, zero3, zero4, zero5, zero6;
+ struct mali_draw_packed postfix;
} __attribute__((packed));
/* Purposeful off-by-one in width, height fields. For example, a (64, 64)
#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 */
projects / mesa.git / blobdiff