hrx-rs 0.7.0

Owned GPU and NPU execution with Loom and IRON compilation
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218

use super::*;

/// Numerical contracts for RGB model inputs and outputs.
#[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
pub enum RgbEncoding {
    /// Encode byte / 255; decode float * 255 without clipping or rounding.
    Unit,
    /// Encode byte / 127.5 - 1; decode round-even((clamp(x,-1,1)+1)*127.5).
    Symmetric,
}

impl ImageOps {
    /// Split packed U8 NHWC into interleaved F32 NCHW views. A factor of f
    /// produces f² views per image: view (dy,dx) samples image[y*f+dy,x*f+dx].
    /// This is a lossless layout permutation before the selected normalization,
    /// not a resize. Both spatial dimensions must divide the factor.
    pub fn prepare_encode_rgb_views(
        &self,
        input: &TensorDesc,
        factor: usize,
        encoding: RgbEncoding,
    ) -> Result<Arc<PreparedModel>> {
        if input.dtype() != DType::U8 || input.layout() != Layout::Nhwc {
            return Err(Error::Message("RGB view encoding requires U8 NHWC".into()));
        }
        self.prepare_rgb_views(input, factor, encoding)
    }

    /// Reassemble interleaved F32 NCHW views as F32 NHWC byte-level RGB.
    /// The view batch must be a multiple of factor². Unit values retain the
    /// network's fractional output; Symmetric values use the documented clamp.
    pub fn prepare_decode_rgb_views(
        &self,
        input: &TensorDesc,
        factor: usize,
        encoding: RgbEncoding,
    ) -> Result<Arc<PreparedModel>> {
        if input.dtype() != DType::F32 || input.layout() != Layout::Nchw {
            return Err(Error::Message("RGB view decoding requires F32 NCHW".into()));
        }
        self.prepare_rgb_views(input, factor, encoding)
    }

    fn prepare_rgb_views(
        &self,
        input: &TensorDesc,
        factor: usize,
        encoding: RgbEncoding,
    ) -> Result<Arc<PreparedModel>> {
        if !input.is_contiguous()
            || input.is_empty()
            || input.shape().len() != 4
            || input.elements() > i32::MAX as usize
            || factor == 0
        {
            return Err(Error::Message("invalid RGB view dimensions".into()));
        }
        let ff = factor
            .checked_mul(factor)
            .ok_or_else(|| Error::Message("RGB view factor overflow".into()))?;
        let encode = input.dtype() == DType::U8;
        let s = input.shape();
        let (batch, height, width) = if encode {
            if s[3] != 3 || !s[1].is_multiple_of(factor) || !s[2].is_multiple_of(factor) {
                return Err(Error::Message(
                    "RGB axes must divide the view factor".into(),
                ));
            }
            (s[0], s[1], s[2])
        } else {
            if s[1] != 3 || !s[0].is_multiple_of(ff) {
                return Err(Error::Message(
                    "RGB view batch must divide factor squared".into(),
                ));
            }
            (s[0] / ff, s[2] * factor, s[3] * factor)
        };
        let vh = height / factor;
        let vw = width / factor;
        let output = if encode {
            TensorDesc::new(DType::F32, vec![batch * ff, 3, vh, vw])?.with_layout(Layout::Nchw)?
        } else {
            TensorDesc::new(DType::F32, vec![batch, height, width, 3])?.with_layout(Layout::Nhwc)?
        };
        self.view_plans.get_or_prepare((input.clone(), factor, encoding), || {
            let count = input.elements();
            let address = if encode {
                format!(r#"
    %x = index.rem %i, %vw : index
    %yy = index.div %i, %vw : index
    %y = index.rem %yy, %vh : index
    %cc = index.div %i, %plane : index
    %ch = index.rem %cc, %three : index
    %view = index.div %cc, %three : index
    %image = index.div %view, %ff : index
    %part = index.rem %view, %ff : index
    %dx = index.rem %part, %factor : index
    %dy = index.div %part, %factor : index
    %xf = index.mul %x, %factor : index
    %xx = index.add %xf, %dx : index
    %yf = index.mul %y, %factor : index
    %yp = index.add %yf, %dy : index
    %ih = index.mul %image, %height : index
    %row = index.add %ih, %yp : index
    %rw = index.mul %row, %width : index
    %pixel = index.add %rw, %xx : index
    %base = index.mul %pixel, %three : index
    %addr = index.add %base, %ch : index
    %at = index.assume %addr [range(%addr, 0, {count})] : index
    %byte = view.load %src[%at] : view<{count}xi8> -> i8
    %uint = scalar.extui %byte : i8 to i32
    %value = scalar.uitofp %uint : i32 to f32
"#)
            } else {
                format!(r#"
    %ch = index.rem %i, %three : index
    %pixel = index.div %i, %three : index
    %x = index.rem %pixel, %width : index
    %yy = index.div %pixel, %width : index
    %y = index.rem %yy, %height : index
    %image = index.div %yy, %height : index
    %dx = index.rem %x, %factor : index
    %dy = index.rem %y, %factor : index
    %xx = index.div %x, %factor : index
    %yp = index.div %y, %factor : index
    %df = index.mul %dy, %factor : index
    %part = index.add %df, %dx : index
    %im = index.mul %image, %ff : index
    %view = index.add %im, %part : index
    %vc = index.mul %view, %three : index
    %channel = index.add %vc, %ch : index
    %hh = index.mul %channel, %vh : index
    %row = index.add %hh, %yp : index
    %rw = index.mul %row, %vw : index
    %addr = index.add %rw, %xx : index
    %at = index.assume %addr [range(%addr, 0, {count})] : index
    %value = view.load %src[%at] : view<{count}xf32> -> f32
"#)
            };
            let arithmetic = match (encode, encoding) {
                (true, RgbEncoding::Unit) => "%scale = scalar.constant 255.0 : f32\n%result = scalar.divf %value, %scale : f32",
                (true, RgbEncoding::Symmetric) => "%scale = scalar.constant 127.5 : f32\n%onef = scalar.constant 1.0 : f32\n%scaled = scalar.divf %value, %scale : f32\n%result = scalar.subf %scaled, %onef : f32",
                (false, RgbEncoding::Unit) => "%scale = scalar.constant 255.0 : f32\n%result = scalar.mulf %value, %scale : f32",
                (false, RgbEncoding::Symmetric) => "%scale = scalar.constant 127.5 : f32\n%onef = scalar.constant 1.0 : f32\n%negative = scalar.constant -1.0 : f32\n%low = scalar.maxnumf %value, %negative : f32\n%clamped = scalar.minnumf %low, %onef : f32\n%shifted = scalar.addf %clamped, %onef : f32\n%scaled = scalar.mulf %shifted, %scale : f32\n%result = scalar.roundevenf %scaled : f32",
            };
            let ty = if encode { "i8" } else { "f32" };
            let source = format!(r#"
amdgpu.target<gfx11-generic> @views_target {{subgroup_size = 32}}
kernel.def target(@views_target) export("rgb_views") @rgb_views(%unused: index) {{
  %one = index.constant 1 : index
  %threads = index.constant 256 : index
  %grid = index.constant {grid} : index
  kernel.launch.config workgroups(%grid, %one, %one) workgroup_size(%threads, %one, %one) : index
}} launch(%unused: index, %input: buffer, %output: buffer) {{
  %off = index.constant 0 : offset
  %threads = index.constant 256 : index
  %lane = kernel.workitem.id<x> : index
  %group = kernel.workgroup.id<x> : index
  %basei = index.mul %group, %threads : index
  %raw = index.add %basei, %lane : index
  %count = index.constant {count} : index
  %three = index.constant 3 : index
  %width = index.constant {width} : index
  %height = index.constant {height} : index
  %vw = index.constant {vw} : index
  %vh = index.constant {vh} : index
  %plane = index.constant {plane} : index
  %factor = index.constant {factor} : index
  %ff = index.constant {ff} : index
  %ig = buffer.assume.memory_space<global> %input : buffer
  %og = buffer.assume.memory_space<global> %output : buffer
  %src = buffer.view %ig[%off] : buffer -> view<{count}x{ty}>
  %dst = buffer.view %og[%off] : buffer -> view<{count}xf32>
  %valid = index.cmp ult, %raw, %count : index
  scf.if %valid {{
    %i = index.assume %raw [range(%raw, 0, {count})] : index
    {address}
    {arithmetic}
    view.store %result, %dst[%i] : f32, view<{count}xf32>
  }}
  kernel.return
}}
"#, grid=count.div_ceil(256), plane=vh*vw);
            let mut model = ModelSession::in_context(&self.context)?;
            let src = model.allocate(input.bytes())?;
            let dst = model.allocate(output.bytes())?;
            // Permutations above are bijections within checked input/output extents.
            let kernel = unsafe { model.compile(&[(&source, Specialization::new("rgb_views"))])? }[0];
            unsafe { model.freeze(&self.context)?.prepare(
                &[Command::Dispatch(Dispatch::indices(kernel, [0], [count.div_ceil(256) as u32,1,1], vec![src.read(),dst.write()]))],
                &[(src,input.clone())], &[(dst,output)], 3,
            ) }
        })
    }

    /// Encode resident pixels without a host boundary.
    pub fn encode_rgb_views(
        &self,
        input: &DeviceTensor,
        factor: usize,
        encoding: RgbEncoding,
    ) -> Result<Inference> {
        self.context.validate(input)?;
        self.prepare_encode_rgb_views(input.desc(), factor, encoding)?
            .submit(std::slice::from_ref(input))
    }
    /// Decode resident model views without a host boundary.
    pub fn decode_rgb_views(
        &self,
        input: &DeviceTensor,
        factor: usize,
        encoding: RgbEncoding,
    ) -> Result<Inference> {
        self.context.validate(input)?;
        self.prepare_decode_rgb_views(input.desc(), factor, encoding)?
            .submit(std::slice::from_ref(input))
    }
}