wip, setting up kit

Author: kali

linalg/src/x86_64_fma/mmm.rs

@@ -3,8 +3,8 @@ use crate::frame::PackedFormat;
 use crate::mmm::MMMKit;
 use crate::mmm::MatMatMulKer;
 use crate::Ops;
-use panel_extract::packed_32_f16_to_f32;
-use panel_extract::packed_32_q40_to_f32;
+use panel_extract::fma_packed_32_f16_to_f32;
+use panel_extract::fma_packed_32_q40_to_f32;
 use tract_data::internal::*;
 use DatumType::*;
 
@@ -63,17 +63,17 @@ pub fn plug(ops: &mut Ops) {
         ops.mmm_kits.push(MMMKit::new_for_mmm(fma_mmm_f32_32x1.mmm(), 1).with_extracting(
             fma_mmm_f32_32x3.mmm(),
             0,
-            packed_32_q40_to_f32.clone(),
+            fma_packed_32_q40_to_f32.clone(),
         ));
         ops.mmm_kits.push(MMMKit::new_for_mmm(fma_mmm_f32_32x1.mmm(), 2).with_extracting(
             fma_mmm_f32_32x3.mmm(),
             1,
-            packed_32_q40_to_f32.clone(),
+            fma_packed_32_q40_to_f32.clone(),
         ));
         ops.mmm_kits.push(
             MMMKit::new(F16, F32, F16, &PackedFormat::new(F16, 32, 32))
                 .with_native(fma_mmm_f32_32x1.mmm(), 3)
-                .with_extracting(fma_mmm_f32_32x3.mmm(), 1, packed_32_f16_to_f32.clone()),
+                .with_extracting(fma_mmm_f32_32x3.mmm(), 1, fma_packed_32_f16_to_f32.clone()),
         );
     }
 }

linalg/src/x86_64_fma/panel_extract.rs

@@ -4,21 +4,25 @@ use crate::Ops;
 use tract_data::internal::*;
 
 pub fn plug(ops: &mut Ops) {
-    ops.panel_extractors.extend([packed_32_q40_to_f32.clone(), packed_32_f16_to_f32.clone()]);
+    ops.panel_extractors.extend([
+        fma_packed_32_q40_to_f32.clone(),
+        fma_packed_32_f16_to_f32.clone(),
+        avx512_packed_128_q40_to_f32.clone(),
+    ]);
 }
 
-panel_extractor!(kernel_packed_32_q40_to_f32 as packed_32_q40_to_f32(
+panel_extractor!(kernel_packed_32_q40_to_f32 as fma_packed_32_q40_to_f32(
     Box::new(super::mmm::pq40_r32()),
     PackedFormat::new(f32::datum_type(), 32, 32)
 ) where(AVX2));
 
-panel_extractor!(kernel_packed_32_f16_to_f32 as packed_32_f16_to_f32(
+panel_extractor!(kernel_packed_32_f16_to_f32 as fma_packed_32_f16_to_f32(
     Box::new(PackedFormat::new(f16::datum_type(), 32, 32)),
     PackedFormat::new(f32::datum_type(), 32, 32)
 ) where(AVX2));
 
-panel_extractor!(kernel_packed_128_q40_to_f32::kernel as packed_128_q40_to_f32(
-    Box::new(super::mmm::PQ40_R128),
+panel_extractor!(kernel_packed_128_q40_to_f32::kernel as avx512_packed_128_q40_to_f32(
+    Box::new(super::mmm::pq40_r128()),
     PackedFormat::new(f32::datum_type(), 128, 32)
 ) where(AVX512F));
 

linalg/x86_64/avx512/avx512_packed_128_q40_to_f32.tmpl

@@ -54,16 +54,13 @@ avx512_packed_128_q40_to_f32_{{suffix}} proc
     push        rdi
     push        rsi
 
+// win: rcx:input rdx: output, r8:k
     mov         rdi, rcx
+    mov         rsi, rdx
+    mov         rdx, r8
 
 {% endif %}
 
-    push        rbx
-    push        r12
-    push        r13
-    push        r14
-    push        r15
-
     sub         rsp, 8
 {% if family == "unix" %}
 .cfi_def_cfa_offset 64
@@ -153,12 +150,6 @@ avx512_packed_128_q40_to_f32_{{suffix}} proc
     ldmxcsr     [rsp + 4]
     add         rsp, 8
 
-    pop r15
-    pop r14
-    pop r13
-    pop r12
-    pop rbx
-
 {% if family == "windows" %}
     pop rsi
     pop rdi