[PATCH v2 01/34] drm/amd/display: fix segment distribution for linear LUTs
Harry Wentland
harry.wentland at amd.com
Fri Sep 8 14:40:07 UTC 2023
On 2023-09-08 10:11, Melissa Wen wrote:
> On 09/06, Harry Wentland wrote:
>> On 2023-08-10 12:02, Melissa Wen wrote:
>>> From: Harry Wentland <harry.wentland at amd.com>
>>>
>>> The region and segment calculation was incapable of dealing
>>> with regions of more than 16 segments. We first fix this.
>>>
>>> Now that we can support regions up to 256 elements we can
>>> define a better segment distribution for near-linear LUTs
>>> for our maximum of 256 HW-supported points.
>>>
>>> With these changes an "identity" LUT looks visually
>>> indistinguishable from bypass and allows us to use
>>> our 3DLUT.
>>>
>>
>> Have you had a chance to test whether this patch makes a
>> difference? I haven't had the time yet.
>
> Last time I tested there was a banding issue on plane shaper LUT PQ ->
> Display Native, but it seems I don't have this use case on tester
> anymore, so I wasn't able to double-check if the issue persist. Maybe
> Joshua can provide some inputs here.
>
> Something I noticed is that shaper LUTs are the only 1D LUT on DCN30
> pipeline that uses cm_helper_translate_curve_to_hw_format(), all others
> (dpp-degamma/dpp-blend/mpc-regamma) call cm3_helper_translate_curve_*.
>
Yeah, they use different codepaths, unfortunately. Might be nice if we
could make them use the same.
> We can drop it from this series until we get the steps to report the
> issue properly.
>
Thanks. If you have concrete steps that show the issue (or even better,
an IGT test) I would be happy to include this.
Harry
> Melissa
>
>>
>> Harry
>>
>>> Signed-off-by: Harry Wentland <harry.wentland at amd.com>
>>> Signed-off-by: Melissa Wen <mwen at igalia.com>
>>> ---
>>> .../amd/display/dc/dcn10/dcn10_cm_common.c | 93 +++++++++++++++----
>>> 1 file changed, 75 insertions(+), 18 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/display/dc/dcn10/dcn10_cm_common.c b/drivers/gpu/drm/amd/display/dc/dcn10/dcn10_cm_common.c
>>> index 3538973bd0c6..04b2e04b68f3 100644
>>> --- a/drivers/gpu/drm/amd/display/dc/dcn10/dcn10_cm_common.c
>>> +++ b/drivers/gpu/drm/amd/display/dc/dcn10/dcn10_cm_common.c
>>> @@ -349,20 +349,37 @@ bool cm_helper_translate_curve_to_hw_format(struct dc_context *ctx,
>>> * segment is from 2^-10 to 2^1
>>> * There are less than 256 points, for optimization
>>> */
>>> - seg_distr[0] = 3;
>>> - seg_distr[1] = 4;
>>> - seg_distr[2] = 4;
>>> - seg_distr[3] = 4;
>>> - seg_distr[4] = 4;
>>> - seg_distr[5] = 4;
>>> - seg_distr[6] = 4;
>>> - seg_distr[7] = 4;
>>> - seg_distr[8] = 4;
>>> - seg_distr[9] = 4;
>>> - seg_distr[10] = 1;
>>> + if (output_tf->tf == TRANSFER_FUNCTION_LINEAR) {
>>> + seg_distr[0] = 0; /* 2 */
>>> + seg_distr[1] = 1; /* 4 */
>>> + seg_distr[2] = 2; /* 4 */
>>> + seg_distr[3] = 3; /* 8 */
>>> + seg_distr[4] = 4; /* 16 */
>>> + seg_distr[5] = 5; /* 32 */
>>> + seg_distr[6] = 6; /* 64 */
>>> + seg_distr[7] = 7; /* 128 */
>>> +
>>> + region_start = -8;
>>> + region_end = 1;
>>> + } else {
>>> + seg_distr[0] = 3; /* 8 */
>>> + seg_distr[1] = 4; /* 16 */
>>> + seg_distr[2] = 4;
>>> + seg_distr[3] = 4;
>>> + seg_distr[4] = 4;
>>> + seg_distr[5] = 4;
>>> + seg_distr[6] = 4;
>>> + seg_distr[7] = 4;
>>> + seg_distr[8] = 4;
>>> + seg_distr[9] = 4;
>>> + seg_distr[10] = 1; /* 2 */
>>> + /* total = 8*16 + 8 + 64 + 2 = */
>>> +
>>> + region_start = -10;
>>> + region_end = 1;
>>> + }
>>> +
>>>
>>> - region_start = -10;
>>> - region_end = 1;
>>> }
>>>
>>> for (i = region_end - region_start; i < MAX_REGIONS_NUMBER ; i++)
>>> @@ -375,16 +392,56 @@ bool cm_helper_translate_curve_to_hw_format(struct dc_context *ctx,
>>>
>>> j = 0;
>>> for (k = 0; k < (region_end - region_start); k++) {
>>> - increment = NUMBER_SW_SEGMENTS / (1 << seg_distr[k]);
>>> + /*
>>> + * We're using an ugly-ish hack here. Our HW allows for
>>> + * 256 segments per region but SW_SEGMENTS is 16.
>>> + * SW_SEGMENTS has some undocumented relationship to
>>> + * the number of points in the tf_pts struct, which
>>> + * is 512, unlike what's suggested TRANSFER_FUNC_POINTS.
>>> + *
>>> + * In order to work past this dilemma we'll scale our
>>> + * increment by (1 << 4) and then do the inverse (1 >> 4)
>>> + * when accessing the elements in tf_pts.
>>> + *
>>> + * TODO: find a better way using SW_SEGMENTS and
>>> + * TRANSFER_FUNC_POINTS definitions
>>> + */
>>> + increment = (NUMBER_SW_SEGMENTS << 4) / (1 << seg_distr[k]);
>>> start_index = (region_start + k + MAX_LOW_POINT) *
>>> NUMBER_SW_SEGMENTS;
>>> - for (i = start_index; i < start_index + NUMBER_SW_SEGMENTS;
>>> + for (i = (start_index << 4); i < (start_index << 4) + (NUMBER_SW_SEGMENTS << 4);
>>> i += increment) {
>>> + struct fixed31_32 in_plus_one, in;
>>> + struct fixed31_32 value, red_value, green_value, blue_value;
>>> + uint32_t t = i & 0xf;
>>> +
>>> if (j == hw_points - 1)
>>> break;
>>> - rgb_resulted[j].red = output_tf->tf_pts.red[i];
>>> - rgb_resulted[j].green = output_tf->tf_pts.green[i];
>>> - rgb_resulted[j].blue = output_tf->tf_pts.blue[i];
>>> +
>>> + in_plus_one = output_tf->tf_pts.red[(i >> 4) + 1];
>>> + in = output_tf->tf_pts.red[i >> 4];
>>> + value = dc_fixpt_sub(in_plus_one, in);
>>> + value = dc_fixpt_shr(dc_fixpt_mul_int(value, t), 4);
>>> + value = dc_fixpt_add(in, value);
>>> + red_value = value;
>>> +
>>> + in_plus_one = output_tf->tf_pts.green[(i >> 4) + 1];
>>> + in = output_tf->tf_pts.green[i >> 4];
>>> + value = dc_fixpt_sub(in_plus_one, in);
>>> + value = dc_fixpt_shr(dc_fixpt_mul_int(value, t), 4);
>>> + value = dc_fixpt_add(in, value);
>>> + green_value = value;
>>> +
>>> + in_plus_one = output_tf->tf_pts.blue[(i >> 4) + 1];
>>> + in = output_tf->tf_pts.blue[i >> 4];
>>> + value = dc_fixpt_sub(in_plus_one, in);
>>> + value = dc_fixpt_shr(dc_fixpt_mul_int(value, t), 4);
>>> + value = dc_fixpt_add(in, value);
>>> + blue_value = value;
>>> +
>>> + rgb_resulted[j].red = red_value;
>>> + rgb_resulted[j].green = green_value;
>>> + rgb_resulted[j].blue = blue_value;
>>> j++;
>>> }
>>> }
>>
More information about the amd-gfx
mailing list