Download build/webgpu/quant-exact-prepare.wgsl.jinja from webgpu-kernels/com.microsoft.QLinearConv: direct link, hf CLI and curl.
- Browser
- Download file 3.58 kB
-
https://huggingface.co/kernels/webgpu-kernels/com.microsoft.QLinearConv/resolve/v1/build/webgpu/quant-exact-prepare.wgsl.jinja
- Command line
-
hf download hf://webgpu-kernels/com.microsoft.QLinearConv@v1/build/webgpu/quant-exact-prepare.wgsl.jinja
-
curl -L -o quant-exact-prepare.wgsl.jinja https://huggingface.co/kernels/webgpu-kernels/com.microsoft.QLinearConv/resolve/v1/build/webgpu/quant-exact-prepare.wgsl.jinja
3.58 kB
| {{ env.wgsl.resourceDeclarations }} | |
| {% if prepareSpatialA or prepareSpatialB %} | |
| {% macro spatial_constants() %} | |
| const KERNEL_H: u32 = {{ kernelHSpec }}u; | |
| const KERNEL_W: u32 = {{ kernelWSpec }}u; | |
| const KERNEL_HW: u32 = KERNEL_H * KERNEL_W; | |
| const KSIZE: u32 = KERNEL_HW; | |
| const STRIDE_H: u32 = {{ strideHSpec }}u; | |
| const STRIDE_W: u32 = {{ strideWSpec }}u; | |
| const DILATION_H: u32 = {{ dilationHSpec }}u; | |
| const DILATION_W: u32 = {{ dilationWSpec }}u; | |
| const PAD_TOP: i32 = {{ padTopSpec }}; | |
| const PAD_LEFT: i32 = {{ padLeftSpec }}; | |
| {% endmacro %} | |
| {% macro spatial_sample(value="value", input="x", zero="x_zero_point", sampleK="k", sampleScalar="auto") %} | |
| let ic = {{ sampleK }} / KSIZE; | |
| let kq = {{ sampleK }} - ic * KSIZE; | |
| let kh = kq / KERNEL_W; | |
| let kw = kq - kh * KERNEL_W; | |
| let oh = pos / params.outW; | |
| let ow = pos - oh * params.outW; | |
| let ih = i32(oh * STRIDE_H + kh * DILATION_H) - PAD_TOP; | |
| let iw = i32(ow * STRIDE_W + kw * DILATION_W) - PAD_LEFT; | |
| var {{ value }}: {{ bScalar if sampleScalar == "auto" else sampleScalar }} = {{ zero }}[0]; | |
| if (ih >= 0 && ih < i32(params.inH) && iw >= 0 && iw < i32(params.inW)) { | |
| {{ value }} = {{ input }}[{% if sampleChannelsLast is defined and sampleChannelsLast %}((batch * params.inH + u32(ih)) * params.inW + u32(iw)) * params.inChannels + ic{% else %}((batch * params.inChannels + ic) * params.inH + u32(ih)) * params.inW + u32(iw){% endif %}]; | |
| } | |
| {% endmacro %} | |
| {{ spatial_constants() }} | |
| {% endif %} | |
| @compute @workgroup_size({{ prepareWorkgroup }}) | |
| fn main(@builtin(workgroup_id) wid: vec3<u32>, @builtin(local_invocation_index) tid: u32) { | |
| let offset = ((wid.y * params.groupsX + wid.x) * {{ prepareWorkgroup }}u + tid) * 4u; | |
| if (offset >= params.count) { return; } | |
| if (offset < params.aCount) { | |
| let m = offset / params.paddedK; | |
| let k = offset % params.paddedK; | |
| {% for h in range(4) %} | |
| { | |
| var value = 0.0; | |
| if (m < params.M && k + {{ h }}u < params.K) { | |
| {% if prepareSpatialA %} | |
| let batch = m / params.outCount; | |
| let pos = m % params.outCount; | |
| let sourceK = k + {{ h }}u; | |
| {{ spatial_sample("sample", "a", "a_zero_point", "sourceK", aScalar) }} | |
| value = f32(i32(sample) - i32(a_zero_point[0])); | |
| {% else %} | |
| value = f32(i32(a[m * params.K + k + {{ h }}u]) - i32(a_zero_point[{{ "m" if prepareZeroPerRow else "0" }}])); | |
| {% endif %} | |
| } | |
| aFloat[offset + {{ h }}u] = {% if precastScalar == "f16" %}f16(value){% else %}value{% endif %}; | |
| } | |
| {% endfor %} | |
| } else { | |
| let bOffset = offset - params.aCount; | |
| {% if prepareBatchB %} | |
| let batch = bOffset / (params.paddedK * params.paddedN); | |
| let localOffset = bOffset % (params.paddedK * params.paddedN); | |
| {% endif %} | |
| let k = {{ "localOffset" if prepareBatchB else "bOffset" }} / params.paddedN; | |
| let n = {{ "localOffset" if prepareBatchB else "bOffset" }} % params.paddedN; | |
| {% for h in range(4) %} | |
| { | |
| var value = 0.0; | |
| if (k < params.K && n + {{ h }}u < params.N) { | |
| {% if prepareSpatialB %} | |
| let pos = n + {{ h }}u; | |
| {{ spatial_sample("sample", "b", "b_zero_point") }} | |
| value = f32(i32(sample) - i32(b_zero_point[0])); | |
| {% else %} | |
| value = f32(i32(b[{% if prepareTransposeB %}(n + {{ h }}u) * params.K + k{% else %}{% if prepareBatchB %}batch * params.K * params.N + {% endif %}k * params.N + n + {{ h }}u{% endif %}]) - i32(b_zero_point[{% if bZeroPerColumn %}n + {{ h }}u{% else %}0{% endif %}])); | |
| {% endif %} | |
| } | |
| bFloat[bOffset + {{ h }}u] = {% if precastScalar == "f16" %}f16(value){% else %}value{% endif %}; | |
| } | |
| {% endfor %} | |
| } | |
| } | |