X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fpanfrost%2Finclude%2Fpanfrost-job.h;h=057aed2419db2c0c82e16e6b33527e8f1a304f93;hb=e5899c1e8818f7cfdd23c06c504009e5659794b7;hp=dfc5d83a80d6136766062c871e1848a0eba92883;hpb=f2f12776248874b2a689cbba8faeb6e4e2144354;p=mesa.git diff --git a/src/panfrost/include/panfrost-job.h b/src/panfrost/include/panfrost-job.h index dfc5d83a80d..057aed2419d 100644 --- a/src/panfrost/include/panfrost-job.h +++ b/src/panfrost/include/panfrost-job.h @@ -30,36 +30,13 @@ #include #include -#include - -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 -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 */ @@ -76,22 +53,13 @@ enum mali_draw_mode { #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 */ @@ -104,41 +72,19 @@ enum mali_func { #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) @@ -201,25 +147,6 @@ struct mali_blend_equation { 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 @@ -229,6 +156,9 @@ struct mali_channel_swizzle { /* 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) @@ -248,11 +178,13 @@ struct mali_channel_swizzle { * for Bifrost framebuffer output. */ #define MALI_FORMAT_SPECIAL2 (7 << 5) +#define MALI_EXTRACT_TYPE(fmt) ((fmt) & 0xe0) /* If the high 3 bits are 3 to 6 these two bits say how many components * 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 @@ -271,113 +203,18 @@ struct mali_channel_swizzle { * MALI_FORMAT_UNORM, it means a 32-bit float. */ #define MALI_CHANNEL_FLOAT 7 +#define MALI_EXTRACT_BITS(fmt) (fmt & 0x7) -enum mali_format { - MALI_RGB565 = MALI_FORMAT_SPECIAL | 0x0, - 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, - - /* YUV formats */ - MALI_NV12 = MALI_FORMAT_SPECIAL | 0xc, - - MALI_Z32_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... */ +/* Applies to midgard1.flags_lo */ -#define MALI_ALPHA_COVERAGE(clampf) ((uint16_t) (int) (clampf * 15.0f)) -#define MALI_GET_ALPHA_COVERAGE(nibble) ((float) nibble / 15.0f) +/* Should be set when the fragment shader updates the depth value. */ +#define MALI_WRITES_Z (1 << 4) -/* Applies to midgard1.flags */ - -/* Should the hardware perform early-Z testing? Normally should be set - * for performance reasons. Clear if you use: discard, - * alpha-to-coverage... * It's also possible this disables - * forward-pixel kill; we're not quite sure which bit is which yet. - * TODO: How does this interact with blending?*/ +/* 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) @@ -391,7 +228,51 @@ enum mali_format { * it might read depth/stencil in particular, also set MALI_READS_ZS */ #define MALI_READS_ZS (1 << 8) -#define MALI_READS_TILEBUFFER (1 << 12) + +/* 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 << 10) + +/* 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 */ @@ -435,6 +316,17 @@ struct midgard_blend_rt { /* On Bifrost systems (all MRT), each render target gets one of these * descriptors */ +enum bifrost_shader_type { + BIFROST_BLEND_F16 = 0, + BIFROST_BLEND_F32 = 1, + BIFROST_BLEND_I32 = 2, + BIFROST_BLEND_U32 = 3, + BIFROST_BLEND_I16 = 4, + BIFROST_BLEND_U16 = 5, +}; + +#define BIFROST_MAX_RENDER_TARGET_COUNT 8 + struct bifrost_blend_rt { /* This is likely an analogue of the flags on * midgard_blend_rt */ @@ -451,57 +343,53 @@ struct bifrost_blend_rt { * constant_hi = int(f / 255) * constant_lo = 65535*f - (65535/255) * constant_hi */ - u16 constant; struct mali_blend_equation equation; + /* * - 0x19 normally * - 0x3 when this slot is unused (everything else is 0 except the index) * - 0x11 when this is the fourth slot (and it's used) -+ * - 0 when there is a blend shader + * - 0 when there is a blend shader */ u16 unk2; + /* increments from 0 to 3 */ u16 index; - union { - struct { - /* So far, I've only seen: - * - R001 for 1-component formats - * - RG01 for 2-component formats - * - RGB1 for 3-component formats - * - RGBA for 4-component formats - */ - u32 swizzle : 12; - enum mali_format format : 8; - - /* Type of the shader output variable. Note, this can - * be different from the format. - * - * 0: f16 (mediump float) - * 1: f32 (highp float) - * 2: i32 (highp int) - * 3: u32 (highp uint) - * 4: i16 (mediump int) - * 5: u16 (mediump uint) - */ - u32 shader_type : 3; - u32 zero : 9; - }; - - /* Only the low 32 bits of the blend shader are stored, the - * high 32 bits are implicitly the same as the original shader. - * According to the kernel driver, the program counter for - * shaders is actually only 24 bits, so shaders cannot cross - * the 2^24-byte boundary, and neither can the blend shader. - * The blob handles this by allocating a 2^24 byte pool for - * shaders, and making sure that any blend shaders are stored - * in the same pool as the original shader. The kernel will - * make sure this allocation is aligned to 2^24 bytes. - */ - u32 shader; - }; + union { + struct { + /* So far, I've only seen: + * - R001 for 1-component formats + * - RG01 for 2-component formats + * - RGB1 for 3-component formats + * - RGBA for 4-component formats + */ + u32 swizzle : 12; + enum mali_format format : 8; + + /* Type of the shader output variable. Note, this can + * be different from the format. + * enum bifrost_shader_type + */ + u32 zero1 : 4; + u32 shader_type : 3; + u32 zero2 : 5; + }; + + /* Only the low 32 bits of the blend shader are stored, the + * high 32 bits are implicitly the same as the original shader. + * According to the kernel driver, the program counter for + * shaders is actually only 24 bits, so shaders cannot cross + * the 2^24-byte boundary, and neither can the blend shader. + * The blob handles this by allocating a 2^24 byte pool for + * shaders, and making sure that any blend shaders are stored + * in the same pool as the original shader. The kernel will + * make sure this allocation is aligned to 2^24 bytes. + */ + u32 shader; + }; } __attribute__((packed)); /* Descriptor for the shader. Following this is at least one, up to four blend @@ -521,12 +409,12 @@ struct mali_shader_meta { } bifrost1; struct { unsigned uniform_buffer_count : 4; - unsigned flags : 12; + unsigned flags_lo : 12; /* vec4 units */ unsigned work_count : 5; unsigned uniform_count : 5; - unsigned unknown2 : 6; + unsigned flags_hi : 6; } midgard1; }; @@ -536,15 +424,19 @@ struct mali_shader_meta { 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 { @@ -580,6 +472,11 @@ struct mali_shader_meta { * - 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 @@ -593,8 +490,7 @@ struct mali_shader_meta { } midgard2; }; - /* zero on bifrost */ - u32 unknown2_8; + u32 padding; /* Blending information for the older non-MRT Midgard HW. Check for * MALI_HAS_BLEND_SHADER to decide how to interpret. @@ -624,21 +520,6 @@ struct mali_job_descriptor_header { 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 */ @@ -772,82 +653,8 @@ struct mali_payload_write_value { * 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 */ @@ -856,21 +663,6 @@ struct mali_attr_meta { /* ORed into an MFBD address to specify the fbx section is included */ #define MALI_MFBD_TAG_EXTRA (0x2) -struct mali_uniform_buffer_meta { - /* This 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. - */ - u64 size : 10; - - /* This 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. - */ - mali_ptr ptr : 64 - 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. @@ -990,14 +782,6 @@ union midgard_primitive_size { u64 pointer; }; -struct bifrost_vertex_only { - u32 unk2; /* =0x2 */ - - u32 zero0; - - u64 zero1; -} __attribute__((packed)); - struct bifrost_tiler_heap_meta { u32 zero; u32 heap_size; @@ -1007,16 +791,19 @@ struct bifrost_tiler_heap_meta { mali_ptr tiler_heap_end; /* hierarchy weights? but they're still 0 after the job has run... */ - u32 zeros[12]; + u32 zeros[10]; + u32 unk1; + u32 unk7e007e; } __attribute__((packed)); struct bifrost_tiler_meta { - u64 zero0; - u16 hierarchy_mask; + u32 tiler_heap_next_start; /* To be written by the GPU */ + u32 used_hierarchy_mask; /* To be written by the GPU */ + u16 hierarchy_mask; /* Five values observed: 0xa, 0x14, 0x28, 0x50, 0xa0 */ u16 flags; u16 width; u16 height; - u64 zero1; + u64 zero0; mali_ptr tiler_heap_meta; /* TODO what is this used for? */ u64 zeros[20]; @@ -1029,24 +816,25 @@ struct bifrost_tiler_only { mali_ptr tiler_meta; u64 zero1, zero2, zero3, zero4, zero5, zero6; - - u32 gl_enables; - u32 zero7; - 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 { + 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. */ @@ -1058,11 +846,12 @@ struct mali_vertex_tiler_postfix { */ u64 uniform_buffers; - /* This is a pointer to an array of pointers to the texture + /* 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 texture_trampoline; + 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 @@ -1079,32 +868,14 @@ struct mali_vertex_tiler_postfix { 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 { struct mali_vertex_tiler_prefix prefix; - - u16 gl_enables; // 0x5 - - /* 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; - struct mali_vertex_tiler_postfix postfix; union midgard_primitive_size primitive_size; @@ -1112,7 +883,6 @@ struct midgard_payload_vertex_tiler { struct bifrost_payload_vertex { struct mali_vertex_tiler_prefix prefix; - struct bifrost_vertex_only vertex; struct mali_vertex_tiler_postfix postfix; } __attribute__((packed)); @@ -1127,7 +897,6 @@ struct bifrost_payload_fused { struct bifrost_tiler_only tiler; struct mali_vertex_tiler_postfix tiler_postfix; u64 padding; /* zero */ - struct bifrost_vertex_only vertex; struct mali_vertex_tiler_postfix vertex_postfix; } __attribute__((packed)); @@ -1139,28 +908,6 @@ struct bifrost_payload_fused { #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) @@ -1170,85 +917,6 @@ enum mali_texture_type { /* 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)); - -/* 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. */ @@ -1266,54 +934,6 @@ FIXED_16(float x, bool allow_negative) 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)); - -/* 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. */ @@ -1423,13 +1043,6 @@ struct midgard_tiler_descriptor { 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; @@ -1449,14 +1062,48 @@ struct mali_sfbd_format { unsigned unk3 : 4; }; -struct mali_single_framebuffer { - u32 unknown1; - u32 unknown2; +/* 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)); + +/* Configures multisampling on Bifrost fragment jobs */ + +struct bifrost_multisampling { u64 zero1; - u64 zero0; + u64 zero2; + mali_ptr sample_locations; + u64 zero4; +} __attribute__((packed)); +struct mali_single_framebuffer { + struct mali_shared_memory shared_memory; struct mali_sfbd_format format; u32 clear_flags; @@ -1519,17 +1166,8 @@ struct mali_single_framebuffer { /* 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) -#define MALI_MFBD_FORMAT_SRGB (1 << 2) +#define MALI_MFBD_FORMAT_SRGB (1 << 0) struct mali_rt_format { unsigned unk1 : 32; @@ -1537,9 +1175,11 @@ struct mali_rt_format { unsigned nr_channels : 2; /* MALI_POSITIVE */ - unsigned unk3 : 5; + 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; @@ -1555,7 +1195,14 @@ struct mali_rt_format { unsigned no_preload : 1; } __attribute__((packed)); -struct bifrost_render_target { +/* Flags for afbc.flags and ds_afbc.flags */ + +#define MALI_AFBC_FLAGS 0x10009 + +/* Lossless RGB and RGBA colorspace transform */ +#define MALI_AFBC_YTR (1 << 17) + +struct mali_render_target { struct mali_rt_format format; u64 zero1; @@ -1573,14 +1220,14 @@ struct bifrost_render_target { mali_ptr metadata; u32 stride; // stride in units of tiles - u32 unk; // = 0x20000 + u32 flags; // = 0x20000 } afbc; 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? @@ -1588,7 +1235,7 @@ struct bifrost_render_target { u32 clear_color_4; // always equal, but unclear function? } __attribute__((packed)); -/* An optional part of bifrost_framebuffer. It comes between the main structure +/* An optional part of mali_framebuffer. It comes between the main structure * and the array of render targets. It must be included if any of these are * enabled: * @@ -1597,26 +1244,31 @@ struct bifrost_render_target { * - TODO: Anything else? */ -/* Flags field: note, these are guesses */ +/* flags_hi */ +#define MALI_EXTRA_PRESENT (0x1) -#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 { +struct mali_framebuffer_extra { mali_ptr checksum; /* 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; + + /* 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. */ struct { mali_ptr depth_stencil_afbc_metadata; u32 depth_stencil_afbc_stride; // in units of tiles - u32 zero1; + u32 flags; mali_ptr depth_stencil; @@ -1628,17 +1280,19 @@ struct bifrost_fb_extra { 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; }; - u64 zero3, zero4; + u32 clear_color_1; + u32 clear_color_2; + u64 zero3; } __attribute__((packed)); /* Flags for mfbd_flags */ @@ -1648,27 +1302,23 @@ struct bifrost_fb_extra { #define MALI_MFBD_DEPTH_WRITE (1 << 10) -/* The MFBD contains the extra bifrost_fb_extra section */ +/* The MFBD contains the extra mali_framebuffer_extra section */ #define MALI_MFBD_EXTRA (1 << 13) -struct bifrost_framebuffer { - u32 stack_shift : 4; - u32 unk0 : 28; - - u32 unknown2; // = 0x1f, same as SFBD - mali_ptr scratchpad; +struct mali_framebuffer { + union { + struct mali_shared_memory shared_memory; + struct bifrost_multisampling msaa; + }; - /* 0x10 */ - mali_ptr sample_locations; - mali_ptr unknown1; /* 0x20 */ u16 width1, height1; 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 */ @@ -1676,10 +1326,16 @@ struct bifrost_framebuffer { u32 mfbd_flags : 24; // = 0x100 float clear_depth; - struct midgard_tiler_descriptor tiler; + union { + struct midgard_tiler_descriptor tiler; + struct { + mali_ptr tiler_meta; + u32 zeros[16]; + }; + }; - /* optional: struct bifrost_fb_extra extra */ - /* struct bifrost_render_target rts[] */ + /* optional: struct mali_framebuffer_extra extra */ + /* struct mali_render_target rts[] */ } __attribute__((packed)); #endif /* __PANFROST_JOB_H__ */