Merge pull request #264 from dustletter/wasm-simd

Wasm simd

Author: Andreas Molzer

src/arch/mod.rs

@@ -2,6 +2,7 @@
 
 mod neon;
 mod ssse3;
+mod wasm;
 
 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 use std::is_x86_feature_detected;
@@ -22,6 +23,10 @@ pub fn get_color_convert_line_ycbcr() -> Option<unsafe fn(&[u8], &[u8], &[u8], &
     {
         return Some(neon::color_convert_line_ycbcr);
     }
+    #[cfg(all(target_feature = "simd128", target_arch = "wasm32"))]
+    {
+        return Some(wasm::color_convert_line_ycbcr);
+    }
     #[allow(unreachable_code)]
     None
 }
@@ -41,6 +46,10 @@ pub fn get_dequantize_and_idct_block_8x8(
     {
         return Some(neon::dequantize_and_idct_block_8x8);
     }
+    #[cfg(all(target_feature = "simd128", target_arch = "wasm32"))]
+    {
+        return Some(wasm::dequantize_and_idct_block_8x8);
+    }
     #[allow(unreachable_code)]
     None
 }

src/arch/wasm.rs

@@ -0,0 +1,277 @@
+#[cfg(target_arch = "wasm32")]
+use std::arch::wasm32::*;
+
+#[cfg(target_arch = "wasm32")]
+#[target_feature(enable = "simd128")]
+fn idct8(data: &mut [v128; 8]) {
+    // The fixed-point constants here are obtained by taking the fractional part of the constants
+    // from the non-SIMD implementation and scaling them up by 1<<15. This is because
+    // i16x8_q15mulr_sat(a, b) is effectively equivalent to (a*b)>>15 (except for possibly some
+    // slight differences in rounding).
+
+    // The code here is effectively equivalent to the calls to "kernel" in idct.rs, except that it
+    // doesn't apply any further scaling and fixed point constants have a different precision.
+
+    let p2 = data[2];
+    let p3 = data[6];
+    let p1 = i16x8_q15mulr_sat(i16x8_add_sat(p2, p3), i16x8_splat(17734)); // 0.5411961
+    let t2 = i16x8_sub_sat(
+        i16x8_sub_sat(p1, p3),
+        i16x8_q15mulr_sat(p3, i16x8_splat(27779)), // 0.847759065
+    );
+    let t3 = i16x8_add_sat(p1, i16x8_q15mulr_sat(p2, i16x8_splat(25079))); // 0.765366865
+
+    let p2 = data[0];
+    let p3 = data[4];
+    let t0 = i16x8_add_sat(p2, p3);
+    let t1 = i16x8_sub_sat(p2, p3);
+
+    let x0 = i16x8_add_sat(t0, t3);
+    let x3 = i16x8_sub_sat(t0, t3);
+    let x1 = i16x8_add_sat(t1, t2);
+    let x2 = i16x8_sub_sat(t1, t2);
+
+    let t0 = data[7];
+    let t1 = data[5];
+    let t2 = data[3];
+    let t3 = data[1];
+
+    let p3 = i16x8_add_sat(t0, t2);
+    let p4 = i16x8_add_sat(t1, t3);
+    let p1 = i16x8_add_sat(t0, t3);
+    let p2 = i16x8_add_sat(t1, t2);
+    let p5 = i16x8_add_sat(p3, p4);
+    let p5 = i16x8_add_sat(p5, i16x8_q15mulr_sat(p5, i16x8_splat(5763))); // 0.175875602
+
+    let t0 = i16x8_q15mulr_sat(t0, i16x8_splat(9786)); // 0.298631336
+    let t1 = i16x8_add_sat(
+        i16x8_add_sat(t1, t1),
+        i16x8_q15mulr_sat(t1, i16x8_splat(1741)), // 0.053119869
+    );
+    let t2 = i16x8_add_sat(
+        i16x8_add_sat(t2, i16x8_add_sat(t2, t2)),
+        i16x8_q15mulr_sat(t2, i16x8_splat(2383)), // 0.072711026
+    );
+    let t3 = i16x8_add_sat(t3, i16x8_q15mulr_sat(t3, i16x8_splat(16427))); // 0.501321110
+
+    let p1 = i16x8_sub_sat(p5, i16x8_q15mulr_sat(p1, i16x8_splat(29490))); // 0.899976223
+    let p2 = i16x8_sub_sat(
+        i16x8_sub_sat(i16x8_sub_sat(p5, p2), p2),
+        i16x8_q15mulr_sat(p2, i16x8_splat(18446)), // 0.562915447
+    );
+
+    let p3 = i16x8_sub_sat(
+        i16x8_q15mulr_sat(p3, i16x8_splat(-31509)), // -0.961570560
+        p3,
+    );
+    let p4 = i16x8_q15mulr_sat(p4, i16x8_splat(-12785)); // -0.390180644
+
+    let t3 = i16x8_add_sat(i16x8_add_sat(p1, p4), t3);
+    let t2 = i16x8_add_sat(i16x8_add_sat(p2, p3), t2);
+    let t1 = i16x8_add_sat(i16x8_add_sat(p2, p4), t1);
+    let t0 = i16x8_add_sat(i16x8_add_sat(p1, p3), t0);
+
+    data[0] = i16x8_add_sat(x0, t3);
+    data[7] = i16x8_sub_sat(x0, t3);
+    data[1] = i16x8_add_sat(x1, t2);
+    data[6] = i16x8_sub_sat(x1, t2);
+    data[2] = i16x8_add_sat(x2, t1);
+    data[5] = i16x8_sub_sat(x2, t1);
+    data[3] = i16x8_add_sat(x3, t0);
+    data[4] = i16x8_sub_sat(x3, t0);
+}
+
+#[cfg(target_arch = "wasm32")]
+#[target_feature(enable = "simd128")]
+fn transpose8(data: &mut [v128; 8]) {
+    // Transpose a 8x8 matrix with a sequence of interleaving operations.
+    // Naming: dABl contains elements from the *l*ower halves of vectors A and B, interleaved, i.e.
+    // A0 B0 A1 B1 ...
+    // dABCDll contains elements from the lower quarter (ll) of vectors A, B, C, D, interleaved -
+    // A0 B0 C0 D0 A1 B1 C1 D1 ...
+    let d01l = i16x8_shuffle::<0,  8,  1,  9,  2, 10, 3, 11>(data[0], data[1]);
+    let d23l = i16x8_shuffle::<0,  8,  1,  9,  2, 10, 3, 11>(data[2], data[3]);
+    let d45l = i16x8_shuffle::<0,  8,  1,  9,  2, 10, 3, 11>(data[4], data[5]);
+    let d67l = i16x8_shuffle::<0,  8,  1,  9,  2, 10, 3, 11>(data[6], data[7]);
+    let d01h = i16x8_shuffle::<4, 12,  5, 13,  6, 14, 7, 15>(data[0], data[1]);
+    let d23h = i16x8_shuffle::<4, 12,  5, 13,  6, 14, 7, 15>(data[2], data[3]);
+    let d45h = i16x8_shuffle::<4, 12,  5, 13,  6, 14, 7, 15>(data[4], data[5]);
+    let d67h = i16x8_shuffle::<4, 12,  5, 13,  6, 14, 7, 15>(data[6], data[7]);
+
+    // Operating on 32-bits will interleave *consecutive pairs* of 16-bit integers.
+    let d0123ll = i32x4_shuffle::<0, 4, 1, 5>(d01l, d23l);
+    let d0123lh = i32x4_shuffle::<2, 6, 3, 7>(d01l, d23l);
+    let d4567ll = i32x4_shuffle::<0, 4, 1, 5>(d45l, d67l);
+    let d4567lh = i32x4_shuffle::<2, 6, 3, 7>(d45l, d67l);
+    let d0123hl = i32x4_shuffle::<0, 4, 1, 5>(d01h, d23h);
+    let d0123hh = i32x4_shuffle::<2, 6, 3, 7>(d01h, d23h);
+    let d4567hl = i32x4_shuffle::<0, 4, 1, 5>(d45h, d67h);
+    let d4567hh = i32x4_shuffle::<2, 6, 3, 7>(d45h, d67h);
+
+    // Operating on 64-bits will interleave *consecutive quadruples* of 16-bit integers.
+    data[0] = i64x2_shuffle::<0, 2>(d0123ll, d4567ll);
+    data[1] = i64x2_shuffle::<1, 3>(d0123ll, d4567ll);
+    data[2] = i64x2_shuffle::<0, 2>(d0123lh, d4567lh);
+    data[3] = i64x2_shuffle::<1, 3>(d0123lh, d4567lh);
+    data[4] = i64x2_shuffle::<0, 2>(d0123hl, d4567hl);
+    data[5] = i64x2_shuffle::<1, 3>(d0123hl, d4567hl);
+    data[6] = i64x2_shuffle::<0, 2>(d0123hh, d4567hh);
+    data[7] = i64x2_shuffle::<1, 3>(d0123hh, d4567hh);
+}
+
+#[cfg(target_arch = "wasm32")]
+#[target_feature(enable = "simd128")]
+pub fn dequantize_and_idct_block_8x8(
+    coefficients: &[i16; 64],
+    quantization_table: &[u16; 64],
+    output_linestride: usize,
+    output: &mut [u8],
+) {
+    // The loop below will write to positions [output_linestride * i, output_linestride * i + 8)
+    // for 0<=i<8. Thus, the last accessed position is at an offset of output_linestrade * 7 + 7,
+    // and if that position is in-bounds, so are all other accesses.
+    assert!(
+        output.len()
+            > output_linestride
+                .checked_mul(7)
+                .unwrap()
+                .checked_add(7)
+                .unwrap()
+    );
+
+    const SHIFT: u32 = 3;
+
+    // Read the DCT coefficients, scale them up and dequantize them.
+    let mut data = [i16x8_splat(0); 8];
+    unsafe {
+        for i in 0..8 {
+            data[i] = i16x8_shl(
+                i16x8_mul(
+                    v128_load(coefficients.as_ptr().wrapping_add(i * 8) as *const _),
+                    v128_load(quantization_table.as_ptr().wrapping_add(i * 8) as *const _),
+                ),
+                SHIFT,
+            );
+        }
+    }
+
+    // Usual column IDCT - transpose - column IDCT - transpose approach.
+    idct8(&mut data);
+    transpose8(&mut data);
+    idct8(&mut data);
+    transpose8(&mut data);
+
+    for i in 0..8 {
+        // The two passes of the IDCT algorithm give us a factor of 8, so the shift here is
+        // increased by 3.
+        // As values will be stored in a u8, they need to be 128-centered and not 0-centered.
+        // We add 128 with the appropriate shift for that purpose.
+        const OFFSET: i16 = 128 << (SHIFT + 3);
+        // We want rounding right shift, so we should add (1/2) << (SHIFT+3) before shifting.
+        const ROUNDING_BIAS: i16 = (1 << (SHIFT + 3)) >> 1;
+
+        let data_with_offset = i16x8_add_sat(data[i], i16x8_splat(OFFSET + ROUNDING_BIAS));
+
+        // SAFETY: the assert at the start of this function ensures
+        // `output_linestride * i + 7` < output.len(), so all accesses are in-bounds.
+        unsafe {
+            v128_store64_lane::<0>(
+                u8x16_narrow_i16x8(
+                    i16x8_shr(data_with_offset, SHIFT + 3),
+                    i16x8_splat(0),
+                ),
+                output.as_mut_ptr().wrapping_add(output_linestride * i) as *mut _,
+            );
+        }
+    }
+}
+
+#[cfg(target_arch = "wasm32")]
+#[target_feature(enable = "simd128")]
+pub fn color_convert_line_ycbcr(y_slice: &[u8], cb_slice: &[u8], cr_slice: &[u8], output: &mut [u8]) -> usize {
+
+    assert!(output.len() % 3 == 0);
+    let num = output.len() / 3;
+    assert!(num <= y_slice.len());
+    assert!(num <= cb_slice.len());
+    assert!(num <= cr_slice.len());
+
+    let num_vecs = num / 8;
+
+    for i in 0..num_vecs {
+        const SHIFT: u32 = 6;
+        // Load.
+        let y: v128;
+        let cb: v128;
+        let cr: v128;
+        // SAFETY: i is at most `num / 8 - 8`, so the highest v128_load64_zero reads from
+        // [num - 8, num). The above asserts ensure this is in-bounds.
+        unsafe {
+            y = v128_load64_zero(y_slice.as_ptr().wrapping_add(i * 8) as *const _);
+            cb = v128_load64_zero(cb_slice.as_ptr().wrapping_add(i * 8) as *const _);
+            cr = v128_load64_zero(cr_slice.as_ptr().wrapping_add(i * 8) as *const _);
+        }
+
+        // Convert to 16 bit.
+        let y = i16x8_shl(i16x8_extend_low_u8x16(y), SHIFT);
+        let cb = i16x8_shl(i16x8_extend_low_u8x16(cb), SHIFT);
+        let cr = i16x8_shl(i16x8_extend_low_u8x16(cr), SHIFT);
+
+        // Add offsets
+        let c128 = i16x8_splat(128 << SHIFT);
+        let y = i16x8_add_sat(y, i16x8_splat((1 << SHIFT) >> 1));
+        let cb = i16x8_sub_sat(cb, c128);
+        let cr = i16x8_sub_sat(cr, c128);
+
+        // Compute cr * 1.402, cb * 0.34414, cr * 0.71414, cb * 1.772
+        let cr_140200 = i16x8_add_sat(i16x8_q15mulr_sat(cr, i16x8_splat(13173)), cr);
+        let cb_034414 = i16x8_q15mulr_sat(cb, i16x8_splat(11276));
+        let cr_071414 = i16x8_q15mulr_sat(cr, i16x8_splat(23401));
+        let cb_177200 = i16x8_add_sat(i16x8_q15mulr_sat(cb, i16x8_splat(25297)), cb);
+
+        // Last conversion step.
+        let r = i16x8_add_sat(y, cr_140200);
+        let g = i16x8_sub_sat(y, i16x8_add_sat(cb_034414, cr_071414));
+        let b = i16x8_add_sat(y, cb_177200);
+
+        // Shift back and convert to u8.
+        let zero = u8x16_splat(0);
+        let r = u8x16_narrow_i16x8(i16x8_shr(r, SHIFT), zero);
+        let g = u8x16_narrow_i16x8(i16x8_shr(g, SHIFT), zero);
+        let b = u8x16_narrow_i16x8(i16x8_shr(b, SHIFT), zero);
+
+        // Shuffle rrrrrrrrggggggggbbbbbbbb to rgbrgbrgb...
+
+        let rg_lanes = i8x16_shuffle::<0, 16,
+                                       1, 17,
+                                       2, 18,
+                                       3, 19,
+                                       4, 20,
+                                       5, 21,
+                                       6, 22,
+                                       7, 23>(r, g);
+
+        let rgb_low = i8x16_shuffle::<0, 1, 16,         // r0, g0, b0
+                                      2, 3, 17,         // r1, g1, b1
+                                      4, 5, 18,         // r2, g2, b2
+                                      6, 7, 19,         // r3, g3, b3
+                                      8, 9, 20,         // r4, g4, b4
+                                      10>(rg_lanes, b); // r5
+
+        let rgb_hi = i8x16_shuffle::<11, 21, 12,       // g5, b5, r6
+                                     13, 22, 14,       // g6, b6, r7
+                                     15, 23,  0,       // g7, b7, --
+                                      0,  0,  0,       // --, --, --
+                                      0,  0,  0,       // --, --, --
+                                      0>(rg_lanes, b); // --
+
+        // SAFETY: i is at most `output.len() / 24 - 1` so the highest possible write is to
+        // `output.len() - 1`.
+        unsafe {
+            v128_store(output.as_mut_ptr().wrapping_add(24 * i) as *mut _, rgb_low);
+            v128_store64_lane::<0>(rgb_hi, output.as_mut_ptr().wrapping_add(24 * i + 16) as *mut _);
+        }
+    }
+
+    num_vecs * 8
+}

tests/lib.rs

@@ -10,6 +10,7 @@ mod crashtest;
 mod reftest;
 
 #[test]
+#[cfg(all(target_family="wasm", target_os="unknown"))]
 #[wasm_bindgen_test::wasm_bindgen_test]
 fn included_file() {
     const FILE: &[u8] = include_bytes!(concat!(env!("CARGO_MANIFEST_DIR"), "/tests/reftest/images/mozilla/jpg-progressive.jpg"));