Batch Buffer

Batch Buffer — Batchbuffer and blitter support

Functions

Types and Values

#define BATCH_SZ
#define BATCH_RESERVED
struct intel_batchbuffer
#define I915_TILING_Yf
#define I915_TILING_Ys
struct igt_buf

Includes

#include <igt.h>

Description

This library provides some basic support for batchbuffers and using the blitter engine based upon libdrm. A new batchbuffer is allocated with intel_batchbuffer_alloc() and for simple blitter commands submitted with intel_batchbuffer_flush().

It also provides some convenient macros to easily emit commands into batchbuffers. All those macros presume that a pointer to a intel_batchbuffer structure called batch is in scope. The basic macros are BEGIN_BATCH, OUT_BATCH, OUT_RELOC and ADVANCE_BATCH.

Note that this library's header pulls in the i-g-t core library as a dependency.

Functions

intel_batchbuffer_alloc ()

struct intel_batchbuffer *
intel_batchbuffer_alloc (drm_intel_bufmgr *bufmgr,
                         uint32_t devid);

Allocates a new batchbuffer object. devid must be supplied since libdrm doesn't expose it directly.

Parameters

bufmgr

libdrm buffer manager

 

devid

pci device id of the drm device

 

Returns

The allocated and initialized batchbuffer object.


intel_batchbuffer_set_context ()

void
intel_batchbuffer_set_context (struct intel_batchbuffer *batch,
                               drm_intel_context *ctx);

intel_batchbuffer_free ()

void
intel_batchbuffer_free (struct intel_batchbuffer *batch);

Releases all resource of the batchbuffer object batch .

Parameters

batch

batchbuffer object

 

intel_batchbuffer_flush ()

void
intel_batchbuffer_flush (struct intel_batchbuffer *batch);

Submits the batch for execution on the blitter engine, selecting the right ring depending upon the hardware platform.

Parameters

batch

batchbuffer object

 

intel_batchbuffer_flush_on_ring ()

void
intel_batchbuffer_flush_on_ring (struct intel_batchbuffer *batch,
                                 int ring);

Submits the batch for execution on ring .

Parameters

batch

batchbuffer object

 

ring

execbuf ring flag

 

intel_batchbuffer_flush_with_context ()

void
intel_batchbuffer_flush_with_context (struct intel_batchbuffer *batch,
                                      drm_intel_context *context);

Submits the batch for execution on the render engine with the supplied hardware context.

Parameters

batch

batchbuffer object

 

context

libdrm hardware context object

 

intel_batchbuffer_reset ()

void
intel_batchbuffer_reset (struct intel_batchbuffer *batch);

Resets batch by allocating a new gem buffer object as backing storage.

Parameters

batch

batchbuffer object

 

intel_batchbuffer_copy_data ()

uint32_t
intel_batchbuffer_copy_data (struct intel_batchbuffer *batch,
                             const void *data,
                             unsigned int bytes,
                             uint32_t align);

This transfers the given data into the batchbuffer. Note that the length must be DWORD aligned, i.e. multiples of 32bits. The caller must confirm that there is enough space in the batch for the data to be copied.

Parameters

batch

batchbuffer object

 

data

pointer to the data to write into the batchbuffer

 

bytes

number of bytes to write into the batchbuffer

 

align

value in bytes to which we want to align

 

Returns

Offset of copied data.


intel_batchbuffer_emit_reloc ()

void
intel_batchbuffer_emit_reloc (struct intel_batchbuffer *batch,
                              drm_intel_bo *buffer,
                              uint64_t delta,
                              uint32_t read_domains,
                              uint32_t write_domain,
                              int fenced);

Emits both a libdrm relocation entry pointing at buffer and the pre-computed DWORD of batch 's presumed gpu address plus the supplied delta into batch .

Note that fenced is only relevant if buffer is actually tiled.

This is the only way buffers get added to the validate list.

Parameters

batch

batchbuffer object

 

buffer

relocation target libdrm buffer object

 

delta

delta value to add to buffer 's gpu address

 

read_domains

gem domain bits for the relocation

 

write_domain

gem domain bit for the relocation

 

fenced

whether this gpu access requires fences

 

intel_batchbuffer_align ()

uint32_t
intel_batchbuffer_align (struct intel_batchbuffer *batch,
                         uint32_t align);

Aligns the current in-batch offset to the given value.

Parameters

batch

batchbuffer object

 

align

value in bytes to which we want to align

 

Returns

Batchbuffer offset aligned to the given value.


intel_batchbuffer_subdata_alloc ()

void *
intel_batchbuffer_subdata_alloc (struct intel_batchbuffer *batch,
                                 uint32_t size,
                                 uint32_t align);

Verify if sufficient size within batch is available to deny overflow. Then allocate size bytes within batch .

Parameters

batch

batchbuffer object

 

size

amount of bytes need to allocate

 

align

value in bytes to which we want to align

 

Returns

Offset within batch between allocated subdata and base of batch .


intel_batchbuffer_subdata_offset ()

uint32_t
intel_batchbuffer_subdata_offset (struct intel_batchbuffer *batch,
                                  void *ptr);

Parameters

batch

batchbuffer object

 

ptr

pointer to given data

 

Returns

Offset within batch between ptr and base of batch .


intel_batchbuffer_space ()

unsigned int
intel_batchbuffer_space (struct intel_batchbuffer *batch);

intel_batchbuffer_emit_dword ()

void
intel_batchbuffer_emit_dword (struct intel_batchbuffer *batch,
                              uint32_t dword);

intel_batchbuffer_require_space ()

void
intel_batchbuffer_require_space (struct intel_batchbuffer *batch,
                                 unsigned int sz);

BEGIN_BATCH()

#define             BEGIN_BATCH(n, r)

Prepares a batch to emit n DWORDS, flushing it if there's not enough space available.

This macro needs a pointer to an intel_batchbuffer structure called batch in scope.

Parameters

n

number of DWORDS to emit

 

r

number of RELOCS to emit

 

OUT_BATCH()

#define OUT_BATCH(d) intel_batchbuffer_emit_dword(batch, d)

Emits d into a batch.

This macro needs a pointer to an intel_batchbuffer structure called batch in scope.

Parameters

d

DWORD to emit

 

OUT_RELOC_FENCED()

#define             OUT_RELOC_FENCED(buf, read_domains, write_domain, delta)

Emits a fenced relocation into a batch.

This macro needs a pointer to an intel_batchbuffer structure called batch in scope.

Parameters

buf

relocation target libdrm buffer object

 

read_domains

gem domain bits for the relocation

 

write_domain

gem domain bit for the relocation

 

delta

delta value to add to buffer 's gpu address

 

OUT_RELOC()

#define             OUT_RELOC(buf, read_domains, write_domain, delta)

Emits a normal, unfenced relocation into a batch.

This macro needs a pointer to an intel_batchbuffer structure called batch in scope.

Parameters

buf

relocation target libdrm buffer object

 

read_domains

gem domain bits for the relocation

 

write_domain

gem domain bit for the relocation

 

delta

delta value to add to buffer 's gpu address

 

ADVANCE_BATCH

#define             ADVANCE_BATCH()

Completes the batch command emission sequence started with BEGIN_BATCH.

This macro needs a pointer to an intel_batchbuffer structure called batch in scope.


BLIT_COPY_BATCH_START()

#define             BLIT_COPY_BATCH_START(flags)

COLOR_BLIT_COPY_BATCH_START()

#define             COLOR_BLIT_COPY_BATCH_START(flags)

intel_blt_copy ()

void
intel_blt_copy (struct intel_batchbuffer *batch,
                drm_intel_bo *src_bo,
                int src_x1,
                int src_y1,
                int src_pitch,
                drm_intel_bo *dst_bo,
                int dst_x1,
                int dst_y1,
                int dst_pitch,
                int width,
                int height,
                int bpp);

This emits a 2D copy operation using blitter commands into the supplied batch buffer object.

Parameters

batch

batchbuffer object

 

src_bo

source libdrm buffer object

 

src_x1

source pixel x-coordination

 

src_y1

source pixel y-coordination

 

src_pitch

src_bo 's pitch in bytes

 

dst_bo

destination libdrm buffer object

 

dst_x1

destination pixel x-coordination

 

dst_y1

destination pixel y-coordination

 

dst_pitch

dst_bo 's pitch in bytes

 

width

width of the copied rectangle

 

height

height of the copied rectangle

 

bpp

bits per pixel

 

intel_copy_bo ()

void
intel_copy_bo (struct intel_batchbuffer *batch,
               drm_intel_bo *dst_bo,
               drm_intel_bo *src_bo,
               long int size);

This emits a copy operation using blitter commands into the supplied batch buffer object. A total of size bytes from the start of src_bo is copied over to dst_bo . Note that size must be page-aligned.

Parameters

batch

batchbuffer object

 

src_bo

source libdrm buffer object

 

dst_bo

destination libdrm buffer object

 

size

size of the copy range in bytes

 

igt_buf_width ()

unsigned
igt_buf_width (const struct igt_buf *buf);

Computes the width in 32-bit pixels of the given buffer.

Parameters

buf

the i-g-t buffer object

 

Returns

The width of the buffer.


igt_buf_height ()

unsigned
igt_buf_height (const struct igt_buf *buf);

Computes the height in 32-bit pixels of the given buffer.

Parameters

buf

the i-g-t buffer object

 

Returns

The height of the buffer.


igt_blitter_fast_copy ()

void
igt_blitter_fast_copy (struct intel_batchbuffer *batch,
                       const struct igt_buf *src,
                       unsigned  src_delta,
                       unsigned  src_x,
                       unsigned  src_y,
                       unsigned  width,
                       unsigned  height,
                       int bpp,
                       const struct igt_buf *dst,
                       unsigned  dst_delta,
                       unsigned  dst_x,
                       unsigned  dst_y);

Copy src into dst using the gen9 fast copy blitter command.

The source and destination surfaces cannot overlap.

Parameters

batch

batchbuffer object

 

src

source i-g-t buffer object

 

src_delta

offset into the source i-g-t bo

 

src_x

source pixel x-coordination

 

src_y

source pixel y-coordination

 

width

width of the copied rectangle

 

height

height of the copied rectangle

 

dst

destination i-g-t buffer object

 

dst_delta

offset into the destination i-g-t bo

 

dst_x

destination pixel x-coordination

 

dst_y

destination pixel y-coordination

 

igt_blitter_fast_copy__raw ()

void
igt_blitter_fast_copy__raw (int fd,
                            uint32_t src_handle,
                            unsigned int src_delta,
                            unsigned int src_stride,
                            unsigned int src_tiling,
                            unsigned int src_x,
                            unsigned  src_y,
                            unsigned int width,
                            unsigned int height,
                            int bpp,
                            uint32_t dst_handle,
                            unsigned int dst_delta,
                            unsigned int dst_stride,
                            unsigned int dst_tiling,
                            unsigned int dst_x,
                            unsigned  dst_y);

Like igt_blitter_fast_copy(), but talking to the kernel directly.

Parameters

fd

file descriptor of the i915 driver

 

src_handle

GEM handle of the source buffer

 

src_delta

offset into the source GEM bo, in bytes

 

src_stride

Stride (in bytes) of the source buffer

 

src_tiling

Tiling mode of the source buffer

 

src_x

X coordinate of the source region to copy

 

src_y

Y coordinate of the source region to copy

 

width

Width of the region to copy

 

height

Height of the region to copy

 

bpp

source and destination bits per pixel

 

dst_handle

GEM handle of the destination buffer

 

dst_delta

offset into the destination GEM bo, in bytes

 

dst_stride

Stride (in bytes) of the destination buffer

 

dst_tiling

Tiling mode of the destination buffer

 

dst_x

X coordinate of destination

 

dst_y

Y coordinate of destination

 

igt_render_copyfunc_t ()

void
(*igt_render_copyfunc_t) (struct intel_batchbuffer *batch,
                          drm_intel_context *context,
                          const struct igt_buf *src,
                          unsigned  src_x,
                          unsigned  src_y,
                          unsigned  width,
                          unsigned  height,
                          const struct igt_buf *dst,
                          unsigned  dst_x,
                          unsigned  dst_y);

This is the type of the per-platform render copy functions. The platform-specific implementation can be obtained by calling igt_get_render_copyfunc().

A render copy function will emit a batchbuffer to the kernel which executes the specified blit copy operation using the render engine. context is optional and can be NULL.

Parameters

batch

batchbuffer object

 

context

libdrm hardware context to use

 

src

source i-g-t buffer object

 

src_x

source pixel x-coordination

 

src_y

source pixel y-coordination

 

width

width of the copied rectangle

 

height

height of the copied rectangle

 

dst

destination i-g-t buffer object

 

dst_x

destination pixel x-coordination

 

dst_y

destination pixel y-coordination

 

igt_get_render_copyfunc ()

igt_render_copyfunc_t
igt_get_render_copyfunc (int devid);

Parameters

devid

pci device id

 

Returns

The platform-specific render copy function pointer for the device specified with devid . Will return NULL when no render copy function is implemented.


igt_fillfunc_t ()

void
(*igt_fillfunc_t) (struct intel_batchbuffer *batch,
                   const struct igt_buf *dst,
                   unsigned  x,
                   unsigned  y,
                   unsigned  width,
                   unsigned  height,
                   uint8_t color);

This is the type of the per-platform fill functions using media or gpgpu pipeline. The platform-specific implementation can be obtained by calling igt_get_media_fillfunc() or igt_get_gpgpu_fillfunc().

A fill function will emit a batchbuffer to the kernel which executes the specified blit fill operation using the media/gpgpu engine.

Parameters

batch

batchbuffer object

 

dst

destination i-g-t buffer object

 

x

destination pixel x-coordination

 

y

destination pixel y-coordination

 

width

width of the filled rectangle

 

height

height of the filled rectangle

 

color

fill color to use

 

igt_get_media_fillfunc ()

igt_fillfunc_t
igt_get_media_fillfunc (int devid);

Parameters

devid

pci device id

 

Returns

The platform-specific media fill function pointer for the device specified with devid . Will return NULL when no media fill function is implemented.


igt_get_gpgpu_fillfunc ()

igt_fillfunc_t
igt_get_gpgpu_fillfunc (int devid);

Parameters

devid

pci device id

 

Returns

The platform-specific gpgpu fill function pointer for the device specified with devid . Will return NULL when no gpgpu fill function is implemented.


igt_vme_func_t ()

void
(*igt_vme_func_t) (struct intel_batchbuffer *batch,
                   const struct igt_buf *src,
                   unsigned int width,
                   unsigned int height,
                   const struct igt_buf *dst);

igt_get_media_vme_func ()

igt_vme_func_t
igt_get_media_vme_func (int devid);

igt_media_spinfunc_t ()

void
(*igt_media_spinfunc_t) (struct intel_batchbuffer *batch,
                         const struct igt_buf *dst,
                         uint32_t spins);

This is the type of the per-platform media spin functions. The platform-specific implementation can be obtained by calling igt_get_media_spinfunc().

The media spin function emits a batchbuffer for the render engine with the media pipeline selected. The workload consists of a single thread which spins in a tight loop the requested number of times. Each spin increments a counter whose final 32-bit value is written to the destination buffer on completion. This utility provides a simple way to keep the render engine busy for a set time for various tests.

Parameters

batch

batchbuffer object

 

dst

destination i-g-t buffer object

 

spins

number of loops to execute

 

igt_get_media_spinfunc ()

igt_media_spinfunc_t
igt_get_media_spinfunc (int devid);

Parameters

devid

pci device id

 

Returns

The platform-specific media spin function pointer for the device specified with devid . Will return NULL when no media spin function is implemented.

Types and Values

BATCH_SZ

#define BATCH_SZ 4096

BATCH_RESERVED

#define BATCH_RESERVED 16

struct intel_batchbuffer

struct intel_batchbuffer {
	drm_intel_bufmgr *bufmgr;
	uint32_t devid;
	int gen;

	drm_intel_context *ctx;
	drm_intel_bo *bo;

	uint8_t buffer[BATCH_SZ];
	uint8_t *ptr, *end;
};

I915_TILING_Yf

#define I915_TILING_Yf 3

I915_TILING_Ys

#define I915_TILING_Ys 4

struct igt_buf

struct igt_buf {
	drm_intel_bo *bo;
	uint32_t stride;
	uint32_t tiling;
	uint32_t bpp;
	uint32_t *data;
	uint32_t size;
	struct {
		uint32_t offset;
		uint32_t stride;
	} aux;
};

This is a i-g-t buffer object wrapper structure which augments the baseline libdrm buffer object with suitable data needed by the render copy and the fill functions.

Members

drm_intel_bo *bo;

underlying libdrm buffer object

 

uint32_t stride;

stride of the buffer

 

uint32_t tiling;

tiling mode bits

 

uint32_t bpp;

bits per pixel, 8, 16 or 32.

 

uint32_t *data;

pointer to the memory mapping of the buffer

 

uint32_t size;

size of the buffer object