[igt-dev] [RFC v2 4/5] lib/igt_fb: Add support for fp16 formats through conversion
Ville Syrjälä
ville.syrjala at linux.intel.com
Fri Apr 5 19:26:42 UTC 2019
On Tue, Mar 05, 2019 at 06:18:36PM -0800, Kevin Strasser wrote:
> Follow design of P01x conversion to support tests needing pixel data in fp16
> (half float 64 bpp).
>
> rfc2:
> - Convert whole rows of pixels if possible (Maarten)
> - Treat rgbx like rgba, let hardware ignore alpha (Maarten)
>
> Signed-off-by: Kevin Strasser <kevin.strasser at intel.com>
lgtm
Reviewed-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> ---
> lib/igt_fb.c | 148 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
> lib/igt_fb.h | 1 +
> 2 files changed, 148 insertions(+), 1 deletion(-)
>
> diff --git a/lib/igt_fb.c b/lib/igt_fb.c
> index 9dca2a4..451b2c2 100644
> --- a/lib/igt_fb.c
> +++ b/lib/igt_fb.c
> @@ -35,6 +35,7 @@
> #include "igt_aux.h"
> #include "igt_color_encoding.h"
> #include "igt_fb.h"
> +#include "igt_halffloat.h"
> #include "igt_kms.h"
> #include "igt_matrix.h"
> #include "igt_vc4.h"
> @@ -161,6 +162,22 @@ static const struct format_desc_struct {
> .num_planes = 1, .plane_bpp = { 32, },
> .hsub = 1, .vsub = 1,
> },
> + { .name = "XRGB16161616F", .depth = -1, .drm_id = DRM_FORMAT_XRGB16161616F,
> + .cairo_id = CAIRO_FORMAT_RGBA128F,
> + .num_planes = 1, .plane_bpp = { 64, },
> + },
> + { .name = "ARGB16161616F", .depth = -1, .drm_id = DRM_FORMAT_ARGB16161616F,
> + .cairo_id = CAIRO_FORMAT_RGBA128F,
> + .num_planes = 1, .plane_bpp = { 64, },
> + },
> + { .name = "XBGR16161616F", .depth = -1, .drm_id = DRM_FORMAT_XBGR16161616F,
> + .cairo_id = CAIRO_FORMAT_RGBA128F,
> + .num_planes = 1, .plane_bpp = { 64, },
> + },
> + { .name = "ABGR16161616F", .depth = -1, .drm_id = DRM_FORMAT_ABGR16161616F,
> + .cairo_id = CAIRO_FORMAT_RGBA128F,
> + .num_planes = 1, .plane_bpp = { 64, },
> + },
> { .name = "NV12", .depth = -1, .drm_id = DRM_FORMAT_NV12,
> .cairo_id = CAIRO_FORMAT_RGB24,
> .num_planes = 2, .plane_bpp = { 8, 16, },
> @@ -661,7 +678,8 @@ static int create_bo_for_fb(struct igt_fb *fb)
> * them, so we need to make sure to use a device BO then.
> */
> if (fb->tiling || fb->size || fb->strides[0] ||
> - (is_i915_device(fd) && igt_format_is_yuv(fb->drm_format)))
> + (is_i915_device(fd) && igt_format_is_yuv(fb->drm_format)) ||
> + (is_i915_device(fd) && igt_format_is_fp16(fb->drm_format)))
> device_bo = true;
>
> /* Sets offets and stride if necessary. */
> @@ -2202,6 +2220,102 @@ static void convert_float_to_yuv16(struct fb_convert *cvt)
> }
> }
>
> +/* { R, G, B, X } */
> +static const unsigned char swizzle_rgbx[] = { 0, 1, 2, 3 };
> +static const unsigned char swizzle_bgrx[] = { 2, 1, 0, 3 };
> +
> +static const unsigned char *rgbx_swizzle(uint32_t format)
> +{
> + switch (format) {
> + default:
> + case DRM_FORMAT_XRGB16161616F:
> + case DRM_FORMAT_ARGB16161616F:
> + return swizzle_bgrx;
> + case DRM_FORMAT_XBGR16161616F:
> + case DRM_FORMAT_ABGR16161616F:
> + return swizzle_rgbx;
> + }
> +}
> +
> +static void convert_fp16_to_float(struct fb_convert *cvt)
> +{
> + int i, j;
> + uint16_t *fp16;
> + float *ptr = cvt->dst.ptr;
> + unsigned int float_stride = cvt->dst.fb->strides[0] / sizeof(*ptr);
> + unsigned int fp16_stride = cvt->src.fb->strides[0] / sizeof(*fp16);
> + const unsigned char *swz = rgbx_swizzle(cvt->src.fb->drm_format);
> + bool needs_reswizzle = swz != swizzle_rgbx;
> +
> + uint16_t *buf = convert_src_get(cvt);
> + fp16 = buf + cvt->src.fb->offsets[0] / sizeof(*buf);
> +
> + for (i = 0; i < cvt->dst.fb->height; i++) {
> + if (needs_reswizzle) {
> + const uint16_t *fp16_tmp = fp16;
> + float *rgb_tmp = ptr;
> +
> + for (j = 0; j < cvt->dst.fb->width; j++) {
> + struct igt_vec4 rgb;
> +
> + igt_half_to_float(fp16_tmp, rgb.d, 4);
> +
> + rgb_tmp[0] = rgb.d[swz[0]];
> + rgb_tmp[1] = rgb.d[swz[1]];
> + rgb_tmp[2] = rgb.d[swz[2]];
> + rgb_tmp[3] = rgb.d[swz[3]];
> +
> + rgb_tmp += 4;
> + fp16_tmp += 4;
> + }
> + } else {
> + igt_half_to_float(fp16, ptr, cvt->dst.fb->width * 4);
> + }
> +
> + ptr += float_stride;
> + fp16 += fp16_stride;
> + }
> +
> + convert_src_put(cvt, buf);
> +}
> +
> +static void convert_float_to_fp16(struct fb_convert *cvt)
> +{
> + int i, j;
> + uint16_t *fp16 = cvt->dst.ptr + cvt->dst.fb->offsets[0];
> + const float *ptr = cvt->src.ptr;
> + unsigned float_stride = cvt->src.fb->strides[0] / sizeof(*ptr);
> + unsigned fp16_stride = cvt->dst.fb->strides[0] / sizeof(*fp16);
> + const unsigned char *swz = rgbx_swizzle(cvt->dst.fb->drm_format);
> + bool needs_reswizzle = swz != swizzle_rgbx;
> +
> + for (i = 0; i < cvt->dst.fb->height; i++) {
> + if (needs_reswizzle) {
> + const float *rgb_tmp = ptr;
> + uint16_t *fp16_tmp = fp16;
> +
> + for (j = 0; j < cvt->dst.fb->width; j++) {
> + struct igt_vec4 rgb;
> +
> + rgb.d[0] = rgb_tmp[swz[0]];
> + rgb.d[1] = rgb_tmp[swz[1]];
> + rgb.d[2] = rgb_tmp[swz[2]];
> + rgb.d[3] = rgb_tmp[swz[3]];
> +
> + igt_float_to_half(rgb.d, fp16_tmp, 4);
> +
> + rgb_tmp += 4;
> + fp16_tmp += 4;
> + }
> + } else {
> + igt_float_to_half(ptr, fp16, cvt->dst.fb->width * 4);
> + }
> +
> + ptr += float_stride;
> + fp16 += fp16_stride;
> + }
> +}
> +
> static void convert_pixman(struct fb_convert *cvt)
> {
> pixman_format_code_t src_pixman = drm_format_to_pixman(cvt->src.fb->drm_format);
> @@ -2290,6 +2404,12 @@ static void fb_convert(struct fb_convert *cvt)
> case DRM_FORMAT_P016:
> convert_yuv16_to_float(cvt);
> return;
> + case DRM_FORMAT_XRGB16161616F:
> + case DRM_FORMAT_XBGR16161616F:
> + case DRM_FORMAT_ARGB16161616F:
> + case DRM_FORMAT_ABGR16161616F:
> + convert_fp16_to_float(cvt);
> + return;
> }
> } else if (cvt->src.fb->drm_format == IGT_FORMAT_FLOAT) {
> switch (cvt->dst.fb->drm_format) {
> @@ -2298,6 +2418,12 @@ static void fb_convert(struct fb_convert *cvt)
> case DRM_FORMAT_P016:
> convert_float_to_yuv16(cvt);
> return;
> + case DRM_FORMAT_XRGB16161616F:
> + case DRM_FORMAT_XBGR16161616F:
> + case DRM_FORMAT_ARGB16161616F:
> + case DRM_FORMAT_ABGR16161616F:
> + convert_float_to_fp16(cvt);
> + return;
> }
> }
>
> @@ -2453,6 +2579,7 @@ cairo_surface_t *igt_get_cairo_surface(int fd, struct igt_fb *fb)
>
> if (fb->cairo_surface == NULL) {
> if (igt_format_is_yuv(fb->drm_format) ||
> + igt_format_is_fp16(fb->drm_format) ||
> ((f->cairo_id == CAIRO_FORMAT_INVALID) &&
> (f->pixman_id != PIXMAN_invalid)))
> create_cairo_surface__convert(fd, fb);
> @@ -2762,6 +2889,25 @@ bool igt_format_is_yuv(uint32_t drm_format)
> }
>
> /**
> + * igt_format_is_fp16
> + * @drm_format: drm fourcc
> + *
> + * Check if the format is fp16.
> + */
> +bool igt_format_is_fp16(uint32_t drm_format)
> +{
> + switch (drm_format) {
> + case DRM_FORMAT_XRGB16161616F:
> + case DRM_FORMAT_ARGB16161616F:
> + case DRM_FORMAT_XBGR16161616F:
> + case DRM_FORMAT_ABGR16161616F:
> + return true;
> + default:
> + return false;
> + }
> +}
> +
> +/**
> * igt_format_plane_bpp:
> * @drm_format: drm fourcc
> * @plane: format plane index
> diff --git a/lib/igt_fb.h b/lib/igt_fb.h
> index e1d885e..07b6814 100644
> --- a/lib/igt_fb.h
> +++ b/lib/igt_fb.h
> @@ -182,6 +182,7 @@ uint32_t igt_drm_format_to_bpp(uint32_t drm_format);
> const char *igt_format_str(uint32_t drm_format);
> bool igt_fb_supported_format(uint32_t drm_format);
> bool igt_format_is_yuv(uint32_t drm_format);
> +bool igt_format_is_fp16(uint32_t drm_format);
> 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);
> --
> 2.7.4
--
Ville Syrjälä
Intel
More information about the igt-dev
mailing list