From 7806d394d645af9d9d830d4b3662d4abef4df347 Mon Sep 17 00:00:00 2001 From: Henry Gressmann Date: Sun, 9 Mar 2025 19:02:35 +0100 Subject: chore: progress on more simd instructions Signed-off-by: Henry Gressmann --- crates/tinywasm/src/interpreter/executor.rs | 107 ++++++++++++++++++++++++-- crates/tinywasm/tests/generated/wasm-simd.csv | 2 +- 2 files changed, 102 insertions(+), 7 deletions(-) diff --git a/crates/tinywasm/src/interpreter/executor.rs b/crates/tinywasm/src/interpreter/executor.rs index aff1696..a00b693 100644 --- a/crates/tinywasm/src/interpreter/executor.rs +++ b/crates/tinywasm/src/interpreter/executor.rs @@ -470,6 +470,13 @@ impl<'store, 'stack> Executor<'store, 'stack> { I32x4MaxU => self.stack.values.calculate_same::(|a, b| Ok(a.simd_max(b))).to_cf()?, I64x2Mul => self.stack.values.calculate_same::(|a, b| Ok(a * b)).to_cf()?, + I16x8Mul => self.stack.values.calculate_same::(|a, b| Ok(a * b)).to_cf()?, + I32x4Mul => self.stack.values.calculate_same::(|a, b| Ok(a * b)).to_cf()?, + + I8x16NarrowI16x8S => unimplemented!(), + I8x16NarrowI16x8U => unimplemented!(), + I16x8NarrowI32x4S => unimplemented!(), + I16x8NarrowI32x4U => unimplemented!(), I8x16AddSatS => self.stack.values.calculate_same::(|a, b| Ok(a.saturating_add(b))).to_cf()?, I16x8AddSatS => self.stack.values.calculate_same::(|a, b| Ok(a.saturating_add(b))).to_cf()?, @@ -480,8 +487,72 @@ impl<'store, 'stack> Executor<'store, 'stack> { I8x16SubSatU => self.stack.values.calculate_same::(|a, b| Ok(a.saturating_sub(b))).to_cf()?, I16x8SubSatU => self.stack.values.calculate_same::(|a, b| Ok(a.saturating_sub(b))).to_cf()?, - I16x8Mul => self.stack.values.calculate_same::(|a, b| Ok(a * b)).to_cf()?, - I32x4Mul => self.stack.values.calculate_same::(|a, b| Ok(a * b)).to_cf()?, + I16x8ExtAddPairwiseI8x16S => unimplemented!(), + I16x8ExtAddPairwiseI8x16U => unimplemented!(), + I32x4ExtAddPairwiseI16x8S => unimplemented!(), + I32x4ExtAddPairwiseI16x8U => unimplemented!(), + + I16x8ExtMulLowI8x16S => unimplemented!(), + I16x8ExtMulLowI8x16U => unimplemented!(), + I16x8ExtMulHighI8x16S => unimplemented!(), + I16x8ExtMulHighI8x16U => unimplemented!(), + I32x4ExtMulLowI16x8S => unimplemented!(), + I32x4ExtMulLowI16x8U => unimplemented!(), + I32x4ExtMulHighI16x8S => unimplemented!(), + I32x4ExtMulHighI16x8U => unimplemented!(), + I64x2ExtMulLowI32x4S => unimplemented!(), + I64x2ExtMulLowI32x4U => unimplemented!(), + I64x2ExtMulHighI32x4S => unimplemented!(), + I64x2ExtMulHighI32x4U => unimplemented!(), + + I16x8ExtendLowI8x16S => unimplemented!(), + I16x8ExtendLowI8x16U => unimplemented!(), + I16x8ExtendHighI8x16S => unimplemented!(), + I16x8ExtendHighI8x16U => unimplemented!(), + I32x4ExtendLowI16x8S => unimplemented!(), + I32x4ExtendLowI16x8U => unimplemented!(), + I32x4ExtendHighI16x8S => unimplemented!(), + I32x4ExtendHighI16x8U => unimplemented!(), + I64x2ExtendLowI32x4S => unimplemented!(), + I64x2ExtendLowI32x4U => unimplemented!(), + I64x2ExtendHighI32x4S => unimplemented!(), + I64x2ExtendHighI32x4U => unimplemented!(), + + I8x16Popcnt => self.stack.values.replace_top::(|v| Ok(v.count_ones())).to_cf()?, + + I16x8Q15MulrSatS => self.stack.values.calculate_same::(|a, b| { + let subq15mulr = |a,b| { + let a = a as i32; + let b = b as i32; + let r = (a * b + 0x4000) >> 15; + if r > i16::MAX as i32 { + i16::MAX + } else if r < i16::MIN as i32 { + i16::MIN + } else { + r as i16 + } + }; + Ok(Simd::::from_array([ + subq15mulr(a[0], b[0]), + subq15mulr(a[1], b[1]), + subq15mulr(a[2], b[2]), + subq15mulr(a[3], b[3]), + subq15mulr(a[4], b[4]), + subq15mulr(a[5], b[5]), + subq15mulr(a[6], b[6]), + subq15mulr(a[7], b[7]), + ])) + }).to_cf()?, + + I32x4DotI16x8S => self.stack.values.calculate::(|a, b| { + Ok(Simd::::from_array([ + i32::from(a[0] * b[0] + a[1] * b[1]), + i32::from(a[2] * b[2] + a[3] * b[3]), + i32::from(a[4] * b[4] + a[5] * b[5]), + i32::from(a[6] * b[6] + a[7] * b[7]), + ])) + }).to_cf()?, F32x4Ceil => self.stack.values.replace_top_same::(|v| Ok(v.ceil())).to_cf()?, F64x2Ceil => self.stack.values.replace_top_same::(|v| Ok(v.ceil())).to_cf()?, @@ -508,10 +579,34 @@ impl<'store, 'stack> Executor<'store, 'stack> { F32x4Max => self.stack.values.calculate_same::(|a, b| Ok(a.simd_max(b))).to_cf()?, F64x2Max => self.stack.values.calculate_same::(|a, b| Ok(a.simd_max(b))).to_cf()?, - F32x4PMin => unimplemented!(), - F32x4PMax => unimplemented!(), - F64x2PMin => unimplemented!(), - F64x2PMax => unimplemented!(), + F32x4PMin => self.stack.values.calculate_same::(|a, b| { + Ok(Simd::::from_array([ + if b[0] < a[0] { b[0] } else { a[0]}, + if b[1] < a[1] { b[1] } else { a[1]}, + if b[2] < a[2] { b[2] } else { a[2]}, + if b[3] < a[3] { b[3] } else { a[3]}, + ])) + }).to_cf()?, + F32x4PMax => self.stack.values.calculate_same::(|a, b| { + Ok(Simd::::from_array([ + if b[0] > a[0] { b[0] } else { a[0]}, + if b[1] > a[1] { b[1] } else { a[1]}, + if b[2] > a[2] { b[2] } else { a[2]}, + if b[3] > a[3] { b[3] } else { a[3]}, + ])) + }).to_cf()?, + F64x2PMin => self.stack.values.calculate_same::(|a, b| { + Ok(Simd::::from_array([ + if b[0] < a[0] { b[0] } else { a[0]}, + if b[1] < a[1] { b[1] } else { a[1]}, + ])) + }).to_cf()?, + F64x2PMax => self.stack.values.calculate_same::(|a, b| { + Ok(Simd::::from_array([ + if b[0] > a[0] { b[0] } else { a[0]}, + if b[1] > a[1] { b[1] } else { a[1]}, + ])) + }).to_cf()?, // not correct I32x4TruncSatF32x4S => self.stack.values.replace_top::(|v| Ok(v.trunc())).to_cf()?, diff --git a/crates/tinywasm/tests/generated/wasm-simd.csv b/crates/tinywasm/tests/generated/wasm-simd.csv index 79adc36..6a25709 100644 --- a/crates/tinywasm/tests/generated/wasm-simd.csv +++ b/crates/tinywasm/tests/generated/wasm-simd.csv @@ -1,2 +1,2 @@ 0.8.0,1300,24679,[{"name":"simd_address.wast","passed":4,"failed":45},{"name":"simd_align.wast","passed":46,"failed":54},{"name":"simd_bit_shift.wast","passed":39,"failed":213},{"name":"simd_bitwise.wast","passed":28,"failed":141},{"name":"simd_boolean.wast","passed":16,"failed":261},{"name":"simd_const.wast","passed":301,"failed":456},{"name":"simd_conversions.wast","passed":48,"failed":234},{"name":"simd_f32x4.wast","passed":16,"failed":774},{"name":"simd_f32x4_arith.wast","passed":16,"failed":1806},{"name":"simd_f32x4_cmp.wast","passed":24,"failed":2583},{"name":"simd_f32x4_pmin_pmax.wast","passed":14,"failed":3873},{"name":"simd_f32x4_rounding.wast","passed":24,"failed":177},{"name":"simd_f64x2.wast","passed":8,"failed":795},{"name":"simd_f64x2_arith.wast","passed":16,"failed":1809},{"name":"simd_f64x2_cmp.wast","passed":24,"failed":2661},{"name":"simd_f64x2_pmin_pmax.wast","passed":14,"failed":3873},{"name":"simd_f64x2_rounding.wast","passed":24,"failed":177},{"name":"simd_i16x8_arith.wast","passed":11,"failed":183},{"name":"simd_i16x8_arith2.wast","passed":19,"failed":153},{"name":"simd_i16x8_cmp.wast","passed":30,"failed":435},{"name":"simd_i16x8_extadd_pairwise_i8x16.wast","passed":4,"failed":17},{"name":"simd_i16x8_extmul_i8x16.wast","passed":12,"failed":105},{"name":"simd_i16x8_q15mulr_sat_s.wast","passed":3,"failed":27},{"name":"simd_i16x8_sat_arith.wast","passed":16,"failed":206},{"name":"simd_i32x4_arith.wast","passed":11,"failed":183},{"name":"simd_i32x4_arith2.wast","passed":26,"failed":123},{"name":"simd_i32x4_cmp.wast","passed":40,"failed":435},{"name":"simd_i32x4_dot_i16x8.wast","passed":3,"failed":27},{"name":"simd_i32x4_extadd_pairwise_i16x8.wast","passed":4,"failed":17},{"name":"simd_i32x4_extmul_i16x8.wast","passed":12,"failed":105},{"name":"simd_i32x4_trunc_sat_f32x4.wast","passed":4,"failed":103},{"name":"simd_i32x4_trunc_sat_f64x2.wast","passed":4,"failed":103},{"name":"simd_i64x2_arith.wast","passed":11,"failed":189},{"name":"simd_i64x2_arith2.wast","passed":2,"failed":23},{"name":"simd_i64x2_cmp.wast","passed":10,"failed":103},{"name":"simd_i64x2_extmul_i32x4.wast","passed":12,"failed":105},{"name":"simd_i8x16_arith.wast","passed":8,"failed":123},{"name":"simd_i8x16_arith2.wast","passed":25,"failed":186},{"name":"simd_i8x16_cmp.wast","passed":30,"failed":415},{"name":"simd_i8x16_sat_arith.wast","passed":24,"failed":190},{"name":"simd_int_to_int_extend.wast","passed":24,"failed":229},{"name":"simd_lane.wast","passed":189,"failed":286},{"name":"simd_linking.wast","passed":0,"failed":3},{"name":"simd_load.wast","passed":8,"failed":31},{"name":"simd_load16_lane.wast","passed":3,"failed":33},{"name":"simd_load32_lane.wast","passed":3,"failed":21},{"name":"simd_load64_lane.wast","passed":3,"failed":13},{"name":"simd_load8_lane.wast","passed":3,"failed":49},{"name":"simd_load_extend.wast","passed":18,"failed":86},{"name":"simd_load_splat.wast","passed":12,"failed":114},{"name":"simd_load_zero.wast","passed":10,"failed":29},{"name":"simd_splat.wast","passed":23,"failed":162},{"name":"simd_store.wast","passed":9,"failed":19},{"name":"simd_store16_lane.wast","passed":3,"failed":33},{"name":"simd_store32_lane.wast","passed":3,"failed":21},{"name":"simd_store64_lane.wast","passed":3,"failed":13},{"name":"simd_store8_lane.wast","passed":3,"failed":49}] -0.9.0-alpha.0,15151,10829,[{"name":"simd_address.wast","passed":49,"failed":0},{"name":"simd_align.wast","passed":100,"failed":0},{"name":"simd_bit_shift.wast","passed":252,"failed":0},{"name":"simd_bitwise.wast","passed":169,"failed":0},{"name":"simd_boolean.wast","passed":277,"failed":0},{"name":"simd_const.wast","passed":757,"failed":0},{"name":"simd_conversions.wast","passed":56,"failed":226},{"name":"simd_f32x4.wast","passed":629,"failed":161},{"name":"simd_f32x4_arith.wast","passed":1357,"failed":465},{"name":"simd_f32x4_cmp.wast","passed":2607,"failed":0},{"name":"simd_f32x4_pmin_pmax.wast","passed":15,"failed":3872},{"name":"simd_f32x4_rounding.wast","passed":148,"failed":53},{"name":"simd_f64x2.wast","passed":639,"failed":164},{"name":"simd_f64x2_arith.wast","passed":1360,"failed":465},{"name":"simd_f64x2_cmp.wast","passed":2685,"failed":0},{"name":"simd_f64x2_pmin_pmax.wast","passed":15,"failed":3872},{"name":"simd_f64x2_rounding.wast","passed":148,"failed":53},{"name":"simd_i16x8_arith.wast","passed":194,"failed":0},{"name":"simd_i16x8_arith2.wast","passed":142,"failed":30},{"name":"simd_i16x8_cmp.wast","passed":436,"failed":29},{"name":"simd_i16x8_extadd_pairwise_i8x16.wast","passed":5,"failed":16},{"name":"simd_i16x8_extmul_i8x16.wast","passed":13,"failed":104},{"name":"simd_i16x8_q15mulr_sat_s.wast","passed":4,"failed":26},{"name":"simd_i16x8_sat_arith.wast","passed":222,"failed":0},{"name":"simd_i32x4_arith.wast","passed":194,"failed":0},{"name":"simd_i32x4_arith2.wast","passed":149,"failed":0},{"name":"simd_i32x4_cmp.wast","passed":450,"failed":25},{"name":"simd_i32x4_dot_i16x8.wast","passed":4,"failed":26},{"name":"simd_i32x4_extadd_pairwise_i16x8.wast","passed":5,"failed":16},{"name":"simd_i32x4_extmul_i16x8.wast","passed":13,"failed":104},{"name":"simd_i32x4_trunc_sat_f32x4.wast","passed":17,"failed":90},{"name":"simd_i32x4_trunc_sat_f64x2.wast","passed":5,"failed":102},{"name":"simd_i64x2_arith.wast","passed":200,"failed":0},{"name":"simd_i64x2_arith2.wast","passed":25,"failed":0},{"name":"simd_i64x2_cmp.wast","passed":97,"failed":16},{"name":"simd_i64x2_extmul_i32x4.wast","passed":13,"failed":104},{"name":"simd_i8x16_arith.wast","passed":131,"failed":0},{"name":"simd_i8x16_arith2.wast","passed":148,"failed":63},{"name":"simd_i8x16_cmp.wast","passed":409,"failed":36},{"name":"simd_i8x16_sat_arith.wast","passed":214,"failed":0},{"name":"simd_int_to_int_extend.wast","passed":25,"failed":228},{"name":"simd_lane.wast","passed":331,"failed":144},{"name":"simd_linking.wast","passed":3,"failed":0},{"name":"simd_load.wast","passed":37,"failed":2},{"name":"simd_load16_lane.wast","passed":36,"failed":0},{"name":"simd_load32_lane.wast","passed":24,"failed":0},{"name":"simd_load64_lane.wast","passed":16,"failed":0},{"name":"simd_load8_lane.wast","passed":52,"failed":0},{"name":"simd_load_extend.wast","passed":20,"failed":84},{"name":"simd_load_splat.wast","passed":14,"failed":112},{"name":"simd_load_zero.wast","passed":12,"failed":27},{"name":"simd_memory-multi.wast","passed":1,"failed":0},{"name":"simd_splat.wast","passed":183,"failed":2},{"name":"simd_store.wast","passed":28,"failed":0},{"name":"simd_store16_lane.wast","passed":4,"failed":32},{"name":"simd_store32_lane.wast","passed":4,"failed":20},{"name":"simd_store64_lane.wast","passed":4,"failed":12},{"name":"simd_store8_lane.wast","passed":4,"failed":48}] +0.9.0-alpha.0,22962,3018,[{"name":"simd_address.wast","passed":49,"failed":0},{"name":"simd_align.wast","passed":100,"failed":0},{"name":"simd_bit_shift.wast","passed":252,"failed":0},{"name":"simd_bitwise.wast","passed":169,"failed":0},{"name":"simd_boolean.wast","passed":277,"failed":0},{"name":"simd_const.wast","passed":757,"failed":0},{"name":"simd_conversions.wast","passed":56,"failed":226},{"name":"simd_f32x4.wast","passed":629,"failed":161},{"name":"simd_f32x4_arith.wast","passed":1357,"failed":465},{"name":"simd_f32x4_cmp.wast","passed":2607,"failed":0},{"name":"simd_f32x4_pmin_pmax.wast","passed":3887,"failed":0},{"name":"simd_f32x4_rounding.wast","passed":148,"failed":53},{"name":"simd_f64x2.wast","passed":639,"failed":164},{"name":"simd_f64x2_arith.wast","passed":1360,"failed":465},{"name":"simd_f64x2_cmp.wast","passed":2685,"failed":0},{"name":"simd_f64x2_pmin_pmax.wast","passed":3887,"failed":0},{"name":"simd_f64x2_rounding.wast","passed":148,"failed":53},{"name":"simd_i16x8_arith.wast","passed":194,"failed":0},{"name":"simd_i16x8_arith2.wast","passed":142,"failed":30},{"name":"simd_i16x8_cmp.wast","passed":436,"failed":29},{"name":"simd_i16x8_extadd_pairwise_i8x16.wast","passed":5,"failed":16},{"name":"simd_i16x8_extmul_i8x16.wast","passed":13,"failed":104},{"name":"simd_i16x8_q15mulr_sat_s.wast","passed":30,"failed":0},{"name":"simd_i16x8_sat_arith.wast","passed":222,"failed":0},{"name":"simd_i32x4_arith.wast","passed":194,"failed":0},{"name":"simd_i32x4_arith2.wast","passed":149,"failed":0},{"name":"simd_i32x4_cmp.wast","passed":450,"failed":25},{"name":"simd_i32x4_dot_i16x8.wast","passed":14,"failed":16},{"name":"simd_i32x4_extadd_pairwise_i16x8.wast","passed":5,"failed":16},{"name":"simd_i32x4_extmul_i16x8.wast","passed":13,"failed":104},{"name":"simd_i32x4_trunc_sat_f32x4.wast","passed":17,"failed":90},{"name":"simd_i32x4_trunc_sat_f64x2.wast","passed":5,"failed":102},{"name":"simd_i64x2_arith.wast","passed":200,"failed":0},{"name":"simd_i64x2_arith2.wast","passed":25,"failed":0},{"name":"simd_i64x2_cmp.wast","passed":97,"failed":16},{"name":"simd_i64x2_extmul_i32x4.wast","passed":13,"failed":104},{"name":"simd_i8x16_arith.wast","passed":131,"failed":0},{"name":"simd_i8x16_arith2.wast","passed":179,"failed":32},{"name":"simd_i8x16_cmp.wast","passed":409,"failed":36},{"name":"simd_i8x16_sat_arith.wast","passed":214,"failed":0},{"name":"simd_int_to_int_extend.wast","passed":25,"failed":228},{"name":"simd_lane.wast","passed":331,"failed":144},{"name":"simd_linking.wast","passed":3,"failed":0},{"name":"simd_load.wast","passed":37,"failed":2},{"name":"simd_load16_lane.wast","passed":36,"failed":0},{"name":"simd_load32_lane.wast","passed":24,"failed":0},{"name":"simd_load64_lane.wast","passed":16,"failed":0},{"name":"simd_load8_lane.wast","passed":52,"failed":0},{"name":"simd_load_extend.wast","passed":20,"failed":84},{"name":"simd_load_splat.wast","passed":14,"failed":112},{"name":"simd_load_zero.wast","passed":12,"failed":27},{"name":"simd_memory-multi.wast","passed":1,"failed":0},{"name":"simd_splat.wast","passed":183,"failed":2},{"name":"simd_store.wast","passed":28,"failed":0},{"name":"simd_store16_lane.wast","passed":4,"failed":32},{"name":"simd_store32_lane.wast","passed":4,"failed":20},{"name":"simd_store64_lane.wast","passed":4,"failed":12},{"name":"simd_store8_lane.wast","passed":4,"failed":48}] -- cgit v1.3.1