[Intel-gfx] [igt-dev] [PATCH V6 i-g-t 3/6] lib: Add function to hash a framebuffer

[Ser, Simon]

Mostly LGTM, here are a few nits.

On Wed, 2019-06-12 at 23:17 -0300, Brian Starkey wrote:
> To use writeback buffers as a CRC source, we need to be able to hash
> them. Implement a simple FVA-1a hashing routine for this purpose.
> 
> Doing a bytewise hash on the framebuffer directly can be very slow if
> the memory is noncached. By making a copy of each line in the FB first
> (which can take advantage of word-access speedup), we can do the hash
> on a cached copy, which is much faster (10x speedup on my platform).
> 
> v6: use igt_memcpy_from_wc() instead of plain memcpy, as suggested by
>     Chris Wilson
> 
> Signed-off-by: Brian Starkey <brian.starkey at arm.com>
> [rebased and updated to the most recent API]
> Signed-off-by: Liviu Dudau <liviu.dudau at arm.com>
> ---
>  lib/igt_fb.c | 66 ++++++++++++++++++++++++++++++++++++++++++++++++++++
>  lib/igt_fb.h |  3 +++
>  2 files changed, 69 insertions(+)
> 
> diff --git a/lib/igt_fb.c b/lib/igt_fb.c
> index 9d4f905e..d07dae39 100644
> --- a/lib/igt_fb.c
> +++ b/lib/igt_fb.c
> @@ -3256,6 +3256,72 @@ bool igt_fb_supported_format(uint32_t drm_format)
>  	return false;
>  }
>  
> +/*
> + * This implements the FNV-1a hashing algorithm instead of CRC, for
> + * simplicity
> + * http://www.isthe.com/chongo/tech/comp/fnv/index.html
> + *
> + * hash = offset_basis
> + * for each octet_of_data to be hashed
> + *         hash = hash xor octet_of_data
> + *         hash = hash * FNV_prime
> + * return hash
> + *
> + * 32 bit offset_basis = 2166136261
> + * 32 bit FNV_prime = 224 + 28 + 0x93 = 16777619
> + */
> +int igt_fb_get_crc(struct igt_fb *fb, igt_crc_t *crc)
> +{
> +#define FNV1a_OFFSET_BIAS 2166136261
> +#define FNV1a_PRIME 16777619

I'd just use plain uint32_t variables for those, but no big deal.

> +	uint32_t hash;
> +	void *map;
> +	char *ptr, *line = NULL;
> +	int x, y, cpp = igt_drm_format_to_bpp(fb->drm_format) / 8;
> +	uint32_t stride = calc_plane_stride(fb, 0);

We could return -EINVAL in case fb->num_planes != 1.

> +	if (fb->is_dumb)
> +		map = kmstest_dumb_map_buffer(fb->fd, fb->gem_handle, fb->size,
> +					      PROT_READ);
> +	else
> +		map = gem_mmap__gtt(fb->fd, fb->gem_handle, fb->size,
> +				    PROT_READ);
> +	ptr = map;

Nit: no need for this, can assign the result of mmap directly to ptr.

> +
> +	/*
> +	 * Framebuffers are often uncached, which can make byte-wise accesses
> +	 * very slow. We copy each line of the FB into a local buffer to speed
> +	 * up the hashing.
> +	 */
> +	line = malloc(stride);
> +	if (!line) {
> +		munmap(map, fb->size);
> +		return -ENOMEM;
> +	}
> +
> +	hash = FNV1a_OFFSET_BIAS;
> +
> +	for (y = 0; y < fb->height; y++, ptr += stride) {
> +
> +		igt_memcpy_from_wc(line, ptr, stride);

Nit: no need to copy the whole stride actually, we can just copy
fb->width * cpp since we're only going to read that.

> +
> +		for (x = 0; x < fb->width * cpp; x++) {
> +			hash ^= line[x];
> +			hash *= FNV1a_PRIME;
> +		}
> +	}
> +
> +	crc->n_words = 1;
> +	crc->crc[0] = hash;
> +
> +	free(line);
> +	munmap(map, fb->size);
> +
> +	return 0;
> +#undef FNV1a_OFFSET_BIAS
> +#undef FNV1a_PRIME
> +}
> +
>  /**
>   * igt_format_is_yuv:
>   * @drm_format: drm fourcc
> diff --git a/lib/igt_fb.h b/lib/igt_fb.h
> index adefebe1..a2741c05 100644
> --- a/lib/igt_fb.h
> +++ b/lib/igt_fb.h
> @@ -37,6 +37,7 @@
>  #include <i915_drm.h>
>  
>  #include "igt_color_encoding.h"
> +#include "igt_debugfs.h"
>  
>  /*
>   * Internal format to denote a buffer compatible with pixman's
> @@ -194,5 +195,7 @@ int igt_format_plane_bpp(uint32_t drm_format, int plane);
>  void igt_format_array_fill(uint32_t **formats_array, unsigned int *count,
>  			   bool allow_yuv);
>  
> +int igt_fb_get_crc(struct igt_fb *fb, igt_crc_t *crc);
> +
>  #endif /* __IGT_FB_H__ */
>  
> _______________________________________________
> igt-dev mailing list
> igt-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/igt-dev