| { |
| "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] }, |
| "cases": [ |
| { |
| "name": "nchw_1x64_256_to_512_linear_decoder", |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 256, 256], "dist": "normal", "seed": 731, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_1x64_256_to_512_linear_asymmetric_decoder", |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "asymmetric" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 256, 256], "dist": "normal", "seed": 732, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 512, 512] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] } |
| }, |
| { |
| "name": "nchw_1x3_256_to_512_linear", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 3, 256, 256] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 3, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_1x32_256_to_512_linear_align_corners_large", |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "align_corners" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 32, 256, 256], "dist": "normal", "seed": 751, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_1x3_256_to_512_linear_align_corners_fallback", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "align_corners" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 3, 256, 256] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 3, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_1x3_256_to_512_nearest_2x", |
| "attrs": { "mode": "nearest", "coordinate_transformation_mode": "asymmetric", "nearest_mode": "floor" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 3, 256, 256] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 3, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_1x128_64_to_128_nearest_2x_f16_vae", |
| "preset": "smoke", |
| "attrs": { |
| "mode": "nearest", |
| "coordinate_transformation_mode": "asymmetric", |
| "nearest_mode": "floor", |
| "scales": [1, 1, 2, 2] |
| }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 128, 64, 64], "dist": "normal", "seed": 601, "scale": 0.2 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 128, 128, 128] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 2" }] } |
| }, |
| { |
| "name": "nchw_1x3_176_to_528_nearest_3x", |
| "attrs": { "mode": "nearest", "coordinate_transformation_mode": "asymmetric", "nearest_mode": "floor" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 3, 176, 176] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 3, 528, 528] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_1x3_128_to_512_nearest_4x", |
| "attrs": { |
| "mode": "nearest", |
| "coordinate_transformation_mode": "half_pixel", |
| "nearest_mode": "round_prefer_floor" |
| }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 3, 128, 128] } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 3, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_1x64_256_to_510_linear_w_odd_scalar_cliff", |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 256, 256], "dist": "normal", "seed": 781, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 512, 510] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "nchw_1x64_256_to_512_linear_f16_vec4_decoder", |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel" }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 64, 256, 256], "dist": "normal", "seed": 782, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 2" }] |
| } |
| }, |
| { |
| "name": "nchw_1x64_256_to_510_linear_f16_w_odd_scalar_cliff", |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel" }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 64, 256, 256], "dist": "normal", "seed": 783, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 512, 510] } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 2" }] } |
| }, |
| { |
| "name": "nchw_1x64_512_to_256_linear_downsample_vec4", |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 512, 512], "dist": "normal", "seed": 784, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 256, 256] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "antialias_linear_8x_downsample_tap_explosion_stress", |
| "preset": "stress", |
| "attrs": { |
| "mode": "linear", |
| "antialias": 1, |
| "coordinate_transformation_mode": "half_pixel", |
| "scales": [1, 1, 0.125, 0.125] |
| }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 32, 256, 256], "dist": "normal", "seed": 901, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 32, 32, 32], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] } |
| }, |
| { |
| "name": "generic_nhwc_linear_2x_upsample_scalar_no_vec4_stress", |
| "preset": "stress", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 256, 256, 16], "dist": "normal", "seed": 902, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 512, 16], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] } |
| }, |
| { |
| "name": "generic_rank5_cubic_64tap_unroll_stress", |
| "preset": "stress", |
| "attrs": { "mode": "cubic", "coordinate_transformation_mode": "half_pixel", "cubic_coeff_a": -0.75 }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 1, 32, 32, 32], "dist": "normal", "seed": 903, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 64, 64, 64], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] } |
| }, |
| { |
| "name": "nchw_1x64_256_to_512_linear_opset18_axes_spelling", |
| "provenance": { |
| "notes": "An opset-18 NCHW decoder upsample names the two spatial axes and supplies one 2x scale for each, producing [1,64,512,512]." |
| }, |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel", "axes": [2, 3], "scales": [2, 2] }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 256, 256], "dist": "normal", "seed": 731, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 64, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] |
| } |
| }, |
| { |
| "name": "suffix_linear_f32_c3", |
| "preset": "edge", |
| "attrs": { "mode": "linear" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 64, 3], "dist": "normal", "seed": 993, "scale": 0.3 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 128, 3] } }, |
| "provenance": { |
| "notes": "Aligned and unaligned identity suffixes exercise generic linear interpolation reuse and its scalar fallback." |
| } |
| }, |
| { |
| "name": "suffix_linear_f32_c4", |
| "preset": "edge", |
| "attrs": { "mode": "linear" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 64, 4], "dist": "normal", "seed": 993, "scale": 0.3 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 128, 4] } }, |
| "provenance": { |
| "notes": "Aligned and unaligned identity suffixes exercise generic linear interpolation reuse and its scalar fallback." |
| } |
| }, |
| { |
| "name": "suffix_linear_f32_c7", |
| "preset": "edge", |
| "attrs": { "mode": "linear" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 64, 7], "dist": "normal", "seed": 993, "scale": 0.3 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 128, 7] } }, |
| "provenance": { |
| "notes": "Aligned and unaligned identity suffixes exercise generic linear interpolation reuse and its scalar fallback." |
| } |
| }, |
| { |
| "name": "suffix_linear_f32_c8", |
| "preset": "edge", |
| "attrs": { "mode": "linear" }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 64, 64, 8], "dist": "normal", "seed": 993, "scale": 0.3 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 128, 128, 8] } }, |
| "provenance": { |
| "notes": "Aligned and unaligned identity suffixes exercise generic linear interpolation reuse and its scalar fallback." |
| } |
| }, |
| { |
| "name": "suffix_linear_f16_c8", |
| "preset": "edge", |
| "attrs": { "mode": "linear" }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 128, 128, 8], "dist": "normal", "seed": 993, "scale": 0.3 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 256, 256, 8] } }, |
| "provenance": { |
| "notes": "Aligned and unaligned identity suffixes exercise generic linear interpolation reuse and its scalar fallback." |
| } |
| }, |
| { |
| "name": "suffix_linear_3axes_f32", |
| "preset": "edge", |
| "attrs": { "mode": "linear" }, |
| "inputs": { |
| "x": { "dtype": "float32", "shape": [1, 16, 16, 16, 8], "dist": "normal", "seed": 993, "scale": 0.3 } |
| }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 24, 24, 24, 8] } }, |
| "provenance": { |
| "notes": "Aligned and unaligned identity suffixes exercise generic linear interpolation reuse and its scalar fallback." |
| } |
| }, |
| { |
| "name": "explicit_linear_f16_2x", |
| "preset": "smoke", |
| "attrs": { "mode": "linear", "coordinate_transformation_mode": "half_pixel", "scales": [1, 1, 2, 2] }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 64, 256, 256], "dist": "normal", "seed": 782, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 64, 512, 512] } }, |
| "bench": { |
| "primary": true, |
| "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 2" }] |
| }, |
| "provenance": { |
| "notes": "The f16 decoder benchmark case written with full-rank scales; identical shape and input distribution." |
| } |
| }, |
| { |
| "name": "cubic_suffix_float32_s128_c16_exclude0", |
| "preset": "stress", |
| "attrs": { |
| "mode": "cubic", |
| "coordinate_transformation_mode": "half_pixel", |
| "cubic_coeff_a": -0.75, |
| "exclude_outside": 0 |
| }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 128, 128, 16], "dist": "normal", "seed": 902, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 256, 256, 16], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] }, |
| "provenance": { |
| "notes": "Cubic channel-contiguous interpolation with an aligned unchanged suffix. Metric counts logical input plus output bytes; it does not count all tap reads or describe physical memory traffic." |
| } |
| }, |
| { |
| "name": "cubic_suffix_float16_s128_c16_exclude0", |
| "preset": "stress", |
| "attrs": { |
| "mode": "cubic", |
| "coordinate_transformation_mode": "half_pixel", |
| "cubic_coeff_a": -0.75, |
| "exclude_outside": 0 |
| }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 128, 128, 16], "dist": "normal", "seed": 902, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 256, 256, 16], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 2" }] }, |
| "provenance": { |
| "notes": "Cubic channel-contiguous interpolation with an aligned unchanged suffix. Metric counts logical input plus output bytes; it does not count all tap reads or describe physical memory traffic." |
| } |
| }, |
| { |
| "name": "cubic_suffix_float32_s256_c16_exclude0", |
| "preset": "stress", |
| "attrs": { |
| "mode": "cubic", |
| "coordinate_transformation_mode": "half_pixel", |
| "cubic_coeff_a": -0.75, |
| "exclude_outside": 0 |
| }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 256, 256, 16], "dist": "normal", "seed": 902, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 512, 512, 16], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] }, |
| "provenance": { |
| "notes": "Cubic channel-contiguous interpolation with an aligned unchanged suffix. Metric counts logical input plus output bytes; it does not count all tap reads or describe physical memory traffic." |
| } |
| }, |
| { |
| "name": "cubic_suffix_float16_s256_c16_exclude0", |
| "preset": "stress", |
| "attrs": { |
| "mode": "cubic", |
| "coordinate_transformation_mode": "half_pixel", |
| "cubic_coeff_a": -0.75, |
| "exclude_outside": 0 |
| }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 256, 256, 16], "dist": "normal", "seed": 902, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 512, 512, 16], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 2" }] }, |
| "provenance": { |
| "notes": "Cubic channel-contiguous interpolation with an aligned unchanged suffix. Metric counts logical input plus output bytes; it does not count all tap reads or describe physical memory traffic." |
| } |
| }, |
| { |
| "name": "cubic_suffix_float32_s128_c16_exclude1", |
| "preset": "stress", |
| "attrs": { |
| "mode": "cubic", |
| "coordinate_transformation_mode": "half_pixel", |
| "cubic_coeff_a": -0.75, |
| "exclude_outside": 1 |
| }, |
| "inputs": { "x": { "dtype": "float32", "shape": [1, 128, 128, 16], "dist": "normal", "seed": 902, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float32", "shape": [1, 256, 256, 16], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 4" }] }, |
| "provenance": { |
| "notes": "Cubic channel-contiguous interpolation with an aligned unchanged suffix. Metric counts logical input plus output bytes; it does not count all tap reads or describe physical memory traffic." |
| } |
| }, |
| { |
| "name": "cubic_suffix_float16_s128_c16_exclude1", |
| "preset": "stress", |
| "attrs": { |
| "mode": "cubic", |
| "coordinate_transformation_mode": "half_pixel", |
| "cubic_coeff_a": -0.75, |
| "exclude_outside": 1 |
| }, |
| "inputs": { "x": { "dtype": "float16", "shape": [1, 128, 128, 16], "dist": "normal", "seed": 902, "scale": 1 } }, |
| "outputs": { "y": { "dtype": "float16", "shape": [1, 256, 256, 16], "dist": "empty" } }, |
| "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.x) + numel(shapes.y)) * 2" }] }, |
| "provenance": { |
| "notes": "Cubic channel-contiguous interpolation with an aligned unchanged suffix. Metric counts logical input plus output bytes; it does not count all tap reads or describe physical memory traffic." |
| } |
| } |
| ] |
| } |
|
|
